July 2026 Summaries

A Roboflow workflow for powder-coat inspection uses an RF-DETR Small object-detection model to identify and localize craters, orange peel, paint bubbles, and scratches, then uses Gemini 2.5 Pro to generate a short description of defect type, location, apparent severity, and need for human review. Trained on a focused subset of 2,775 annotated images using a 70/15/15 data split and 384×384 preprocessing, the model achieved 74.0% mAP@50, 72.2% precision, and 77.1% recall, though its performance should be validated under actual production conditions such as varying lighting, gloss, angles, and surface colors. The workflow annotates images with detection boxes and labels, overlays Gemini’s constrained one-sentence summary, and logs original images, predictions, annotations, and summaries through Roboflow Vision Events for later quality analysis. The system can be expanded for continuous camera-based inspections, confidence thresholds, defect counting, notifications, and review routing, but it is intended as an inspection aid because visual detections and severity estimates require human verification rather than serving as final quality decisions.

Grace Freund describes a project where a Raspberry Pi, a phone camera, and a Roboflow classification model are used to monitor a pool's water level and alert when a refill is needed. Inspired by a colleague's demos of Roboflow running on a Raspberry Pi, Freund sets up a system at home in Saint Louis to address her father's need to know the pool's water status without manual checking. The project involves turning a phone into a wireless RTSP camera to stream video to the Raspberry Pi, where the Roboflow Inference server processes the frames locally. After initial setup challenges, such as switching from a GPU-dependent to a CPU-only Roboflow container, Freund captures and labels images of the pool's water level to train a visual state recognition model, specifically using the DINOv3 Small classification model. The Raspberry Pi then serves a live dashboard via a small Flask web app, providing a real-time pool status to household members. Despite the simplicity and affordability of the setup, the project effectively addresses a practical problem with minimal technical expertise required, and future enhancements are considered, such as a permanently mounted camera and automated hose activation.

Mostafa Ibrahim describes a Roboflow Workflow for leather quality inspection that extends RF-DETR defect detection into automated severity decisions based on bounding-box area. Using a public annotated leather-defect dataset, the process involves training an RF-DETR model, detecting defects at a 0.4 confidence threshold, and using a Custom Python block to classify images as PASS when no defects are found, REVIEW when only defects below a configurable 9,000-pixel-square area threshold appear, or FAIL when any defect exceeds that threshold. The workflow also overlays boxes, labels, and inspection status on images, produces a structured JSON quality report, and logs original and annotated images, detections, and metadata to Roboflow Vision Events for traceability. Test examples show large cuts and folds failing automatically, a small but confidently detected cut being routed for review, and clean leather passing without manual intervention. The approach can run on edge hardware such as NVIDIA Jetson or through an API, supports monitoring across production lines and suppliers, and can use production failure cases to improve future model training.

Shop floor data collection records production activity such as completed units, cycle times, defects, downtime, changeovers, machine states, and safety compliance, but conventional methods including paper forms, terminals, barcode scans, and operator-entered MES records can be delayed, incomplete, or lack visual context. The article describes how computer vision can supplement these methods by using cameras, custom RF-DETR models, and Roboflow Workflows to convert visual observations into timestamped, structured events with metadata and evidence images. It explains that reliable vision-based records require event definitions such as tracking a product and registering completion only when it crosses a virtual line, rather than counting repeated detections across video frames. A conveyor example demonstrates how tracked cartons can generate completion events, cycle-time calculations, slow-cycle alerts, JSON logs, annotated video, and evidence images while accounting for issues such as occlusion, counter resets, simultaneous crossings, and camera interruptions. The resulting events can be stored in Vision Events and routed to MES platforms, historians, quality systems, dashboards, spreadsheets, APIs, or PLCs, with edge deployment supporting low-latency decisions and continued operation during connectivity disruptions.

A unified namespace (UNS) is a hierarchical, real-time plant data model, commonly using MQTT, in which systems publish state changes to shared topics and downstream applications subscribe to the information they need, reducing fragile point-to-point integrations. While conventional UNS designs typically include PLCs, SCADA, MES, ERP, and historians, cameras are often excluded because raw video is too large and unstructured for event-based messaging. Vision AI can make cameras first-class industrial sensors by converting video into structured events such as defect detections, production counts, and safety alerts, optionally linking to separately stored images or footage. Effective integration depends on consistent ISA-95-style topic hierarchies and predictable JSON payloads that identify location, camera, event type, confidence, part information, and image references. The post describes using Roboflow Workflows to run inference on camera streams and publish selected detections to MQTT brokers, allowing dashboards, MES platforms, historians, and PLC-facing layers to consume the same events without separate camera-specific integrations. It distinguishes MQTT as the transport mechanism from UNS as the broader shared data model, notes that Sparkplug B is optional, and recommends assessing vendors based on their ability to publish open, structured events into a customer-controlled namespace.

The text outlines a comprehensive guide for automating the detection of empty shelves in retail environments using a computer vision model, RF-DETR, and Roboflow Workflow. It emphasizes the financial impact of out-of-stock items, estimated at $1.2 trillion annually, and the inefficiency of manual shelf monitoring in large stores. The process involves training the RF-DETR model on a dataset of retail shelf images with labeled empty spaces, then using Roboflow Workflow to classify shelves as stocked, partially stocked, or empty based on detection coverage. The workflow provides annotated images indicating restocking needs and logs each scan as a Vision Event, facilitating monitoring over time. The guide also highlights the importance of setting appropriate coverage thresholds tailored to specific retail environments and suggests using Roboflow Agent for easier pipeline assembly. This approach can be adapted for various inventory monitoring tasks by adjusting the model and maintaining the workflow's structure and logic.

Production line monitoring with Camera AI, as detailed in the text, involves transforming traditional cameras into intelligent systems that can detect defects, count products, and trigger actions in real time. By using Roboflow's AI1 Camera and its associated tools, manufacturers can train custom models, like the RF-DETR detection model, to analyze live camera feeds and automate inspection processes, such as bottle inspection on a conveyor belt. This system shifts from passive video recording to active intelligence, allowing manufacturers to detect issues like missing caps or labels on bottles and respond immediately by logging these defects as actionable Vision Events. The process includes capturing data, training models for specific inspection needs, and setting up automated workflows that transform detections into meaningful events. This approach not only enhances quality control by identifying defects early but also improves production efficiency and workplace safety by automating manual inspections and monitoring safety compliance. The text further discusses the advantages of edge versus cloud inference, the importance of connecting vision outputs to real-world systems, and the need for monitoring model performance at scale to ensure consistent quality in dynamic production environments.

The tutorial describes building an automated flange-quality inspection system in Roboflow to detect scratches, cracks, dents, and pinholes before defective parts are installed. It uses a labeled flange dataset from Roboflow Universe to train an RF-DETR Small object-detection model with standard augmentations, then deploys the model in a workflow that translates source labels into English, visualizes detected defects, removes duplicate detections, and assigns PASS, REVIEW, or FAIL outcomes based on confidence thresholds. The workflow produces an annotated image, a structured quality report, and a brief Gemini-generated operator recommendation, while Roboflow Vision Events records inspections with contextual metadata. Examples demonstrate high-confidence defects resulting in failure, lower-confidence detections being routed for manual review, and defect-free images passing automatically. The approach is intended for production use through APIs or edge inference, with reviewed cases added back into training to improve the model and reduce manual inspection over time; Roboflow Agent can alternatively assemble much of the pipeline from a natural-language request.

Google's Gemini 3.6 Flash, released alongside Gemini 3.5 Flash-Lite, is described as a faster and cheaper "workhorse model" that excels in video understanding, surpassing its predecessor, Gemini 3.5 Flash, in most image tasks and offering better cost efficiency. However, it falls short in object detection, often providing inaccurate and loosely defined results compared to its predecessor and the less expensive Flash-Lite. Despite its shortcomings in object detection, Gemini 3.6 Flash leads in tasks like data extraction and counting, indicating that it remains effective in recognizing objects but struggles with precise localization. The model also demonstrates superior performance in video analysis, ranking highest on video benchmarks like VantageBench and VideoNet, with ongoing efforts to enhance video evaluations. While it is recommended for general image understanding and video tasks due to its lower operational costs, for precision-based object detection, users are advised to consider alternative models like a fine-tuned RF-DETR model for more accurate results.

Erik Kokalj's blog post from July 22, 2026, details the process of running RF-DETR within NVIDIA DeepStream on a Jetson Orin NX, explaining both DeepStream and RF-DETR functionalities. DeepStream is described as NVIDIA's video analytics SDK that uses GStreamer to provide hardware-accelerated video processing pipelines, while RF-DETR is a DETR-family detector developed by Roboflow. The post discusses building a TensorRT engine, creating a parser for model output interpretation, and running it on RTSP streams with customized box colors for different classes. The RF-DETR model is used to detect objects by transforming raw data into structured detection results, which involves converting ONNX files into TensorRT engines and configuring DeepStream with a parser function to interpret detection results. The author emphasizes the importance of correctly setting preprocessing parameters and describes how to implement custom per-class color coding using Python, enhancing the visualization of detected objects. Performance metrics indicate that a Jetson Orin NX can efficiently handle a four-stream setup at 30 fps per stream, leveraging the RF-DETR Nano model with FP16 precision on a 384x384 input resolution.

Automating flange quality inspection involves training an RF-DETR model on a labeled dataset of flange defects and integrating it into a Roboflow Workflow, which detects and classifies defects such as cracks, scratches, dents, and pinholes. The system categorizes images as pass, review, or fail based on confidence levels, logging each result to Vision Events and generating operator recommendations through a Gemini block. The process aims to reduce manual inspections over time by feeding review cases back into training, enhancing the model's accuracy. This inspection pipeline not only identifies and visualizes defects but also provides structured reports and quick summaries, leveraging Roboflow's capabilities to streamline quality control in manufacturing environments. The adaptable Workflow can be applied to various inspection tasks by simply swapping datasets and models, while maintaining the same triage and visualization processes.

The tutorial by Aarnav Shah outlines a method for creating automatic highlight reels from kids' soccer games using computer vision technology. It demonstrates a system where two RF-DETR models trained via the Roboflow UI—one for detecting players, balls, and goals, and another for recognizing jersey numbers—are integrated into a Roboflow Workflow with BoT-SORT tracking to identify and follow specific players. A local Python pipeline processes the detections, transforming them into jersey numbers, identifying a single tracked player, detecting goals, and assembling a highlight reel. This system is presented as a prototype, aiming to solve the common issue of capturing valuable moments from lengthy game footage by enabling parents to upload a game, select their child, and receive curated clips. The tutorial includes a FastAPI web app interface for user interaction and emphasizes the iterative process of improving model accuracy through continual training based on feedback from processed videos. While it offers a promising solution for automated sports analytics at a grassroots level, the system acknowledges its current limitations and encourages further refinement through user participation.

A system for automating hog ring inspection in automotive seat assembly is introduced, utilizing Roboflow Workflows, RF-DETR, and Gemini for effective detection and evaluation. The process involves using RF-DETR to detect and count visible hog rings, a Custom Python block to compare detected versus expected counts, and Gemini for a visual review to flag any installation issues. This automation creates a traceable inspection history through Vision Events, facilitating integrations with Manufacturing Execution Systems (MES) and other platforms for real-time alerts and quality checks. The workflow’s adaptability allows for the integration of fixed production-line cameras and can be customized for different seat models by adjusting expected counts. The system improves quality control by detecting potential defects early in the production process, offering a comprehensive solution that can extend to other fasteners beyond hog rings.

Timothy M's blog post, published on July 20, 2026, explores various advanced techniques for optimizing AI inference costs, emphasizing the importance of evaluating the entire computer vision pipeline rather than focusing solely on model speed. It suggests starting with profiling the pipeline to identify the most cost-effective changes that maintain accuracy, such as using a smaller model variant, applying quantization, pruning, or distillation. The post highlights the significance of addressing factors outside the model, especially in video applications, like adjusting inference frame rates, utilizing tracking, and choosing appropriate deployment options. The discussion includes model compression, runtime optimization, and infrastructure strategies, advocating for asynchronous processing, efficient model architectures, and reducing unnecessary operations. It also covers the importance of caching, selecting suitable deployment methods, and ensuring that optimizations are tested incrementally to address specific bottlenecks effectively. The article concludes by encouraging the use of tools like Roboflow Train and Workflows to build scalable and cost-effective computer vision applications.

Cosmetic defect detection in manufacturing settings involves using computer vision to identify visible imperfections such as paint scratches, dents, and chips that, while not affecting product functionality, can impact aesthetics, customer satisfaction, and resale value. In an illustrative guide, Timothy M describes building an end-to-end car-parts inspection system using Roboflow, utilizing a custom RF-DETR model to detect parts like bumpers and doors, and cosmetic defects such as paint scratches. The system captures images of product surfaces, applies object detection to locate defects, and associates them with specific components, generating pass/fail results and repair recommendations. This approach enhances quality control processes across various industries, including automotive, electronics, and furniture, by improving inspection efficiency and consistency, supporting manual reviews, and facilitating repair and quality reporting. Roboflow's platform aids in managing image datasets, training models, and deploying inspection logic to streamline the cosmetic defect detection workflow.

Aarnav Shah presents an innovative approach to preventing package theft by creating a computer vision-based lockbox that automatically secures deliveries. This system utilizes a Roboflow-trained model for real-time person and package detection via a concealed iOS device, with a serverless workflow validating deliveries within a designated porch area. An ESP32 microcontroller, connected to a relay and a solenoid lock, executes unlock commands when certain conditions are met. The design includes a live dashboard for remote monitoring and notifications, ensuring family members can verify deliveries from anywhere. The project, which leverages affordable electronic components and an old iPhone, is detailed for replication on GitHub, emphasizing a preventative strategy over retributive measures like those seen in Mark Rober's glitter bomb series.

The guide outlines the development of an automated beverage shelf-monitoring system using Roboflow Workflows, focusing on detecting Coca-Cola, Fanta, and Sprite bottles on retail shelves. It employs an RF-DETR model to identify and count these bottles, with a Custom Python block assessing the counts against set targets to prioritize restocking. Gemini 2.5 Pro generates inspection summaries from annotated images, enabling staff to quickly identify which shelves need attention. The system addresses common retail challenges such as in-store execution errors that lead to out-of-stock events, by providing a continuous monitoring solution that reduces manual checks and helps prioritize replenishment actions. The workflow's flexibility allows for adjustments in detection, replenishment logic, and summary creation, making it adaptable for broader store deployment through fixed cameras and integration with platforms like Slack for task management.

The advancement of automated tube and pipe defect detection can be achieved by training Roboflow's RF-DETR model on labeled pipe imagery to identify holes, cracks, and ruptures, which are persistent threats to pipeline integrity. This process integrates with a Roboflow Workflow to assign a PASS, REVIEW, or FAIL verdict to each image, with low-confidence detections being directed to human review. The tutorial outlines the workflow, starting with capturing an image that runs through a custom-trained RF-DETR model at a low confidence threshold, followed by a separate filter block that applies the real inspection threshold, ensuring no defects are missed due to threshold settings. The model is trained on a diverse dataset encompassing various damage types under different conditions, and the workflow is designed to adapt to new defect types with additional training data. This approach enhances pipeline inspection reliability by logging each inspection to Vision Events, allowing continuous improvement of the model without altering the workflow structure.

OpenAI's recent release of the GPT-5.6 lineup, including the Sol, Terra, and Luna models, represents a significant advancement in their visual language models (VLMs), focusing on enhancing capabilities in object detection, counting, and document layout understanding. The Sol model, in particular, exhibits substantial improvements over its predecessor, GPT-5.5, notably achieving a higher mean average precision in object detection and improved counting accuracy, though it still faces challenges with large images and complex scenes. While OCR performance remains similar to GPT-5.5, Sol excels in extracting embedded text from intricate visual contexts, despite occasional failures in tasks involving low-contrast or reflective surfaces. Despite these gains, the models require higher token usage, impacting processing costs and latency, making Gemini 3.5 Flash a more cost-effective option for large-scale tasks. Nonetheless, GPT-5.6 marks OpenAI's strengthened focus on vision tasks, positioning it as a competitive choice for screen understanding, document workflows, and visual reasoning applications.

Roboflow's serverless inference API is designed to handle complex machine vision tasks with efficiency by utilizing a unique architecture that separates the process of accepting and executing requests. The system employs a three-layer design with a message broker at its core, enabling stateless Go gateways to manage requests and GPU nodes to execute them based on model availability in VRAM. This structure allows for asynchronous processing while providing a synchronous experience to clients, overcoming challenges like high latency, VRAM multi-tenancy, and asynchronous failures. In production, this architecture supports tens of millions of requests weekly, managing a dynamic catalog of thousands of models with efficient resource allocation and cache management. By ensuring that routing decisions are made at the worker level where the model state is known, the system optimizes performance and reliability, making backpressure structural and failure management robust. This approach allows Roboflow to provide scalable, cost-effective inference services while maintaining high performance and reliability.

Roboflow's Workflows integration with Auto Label allows for the creation of a multi-model consensus pipeline that produces pixel-perfect segmentation masks, leveraging models like SAM 3, Google Gemini, and OpenAI's GPT. This system refines object detections into masks using a rules-based consensus block to retain only those masks that have been independently agreed upon by at least two of the three models. This approach is particularly beneficial in scenarios where the exact shape of an object is crucial, such as in defect analysis and medical imaging, by reducing the time and cost associated with manual annotation. By using a combination of models, users can generate accurate segmentation masks without being constrained to a single model or provider, and the integration is implemented serverlessly for easy deployment on unannotated images. The pipeline ensures high-confidence masks by eliminating uncorrelated errors and trimming away hallucinated spillover, resulting in a reliable dataset ready for human review or direct integration into production datasets.

In the realm of manufacturing, many computer vision pilots fail to progress beyond the pilot phase, not due to technological shortcomings but because of integration challenges and organizational issues. Erik Kokalj highlights in his article that a successful transition from pilot to production requires rethinking computer vision projects as operational changes rather than simple software tasks. Jeff Witt, a Digital Transformation Leader, emphasizes the importance of integrating vision data with existing systems to enhance scalability and usability across multiple sites. Instead of waiting for perfect conditions, he advocates for deploying with current data and improving iteratively. Moreover, by empowering plant teams to manage their own models and use cases, the technology becomes more accessible and effective. The strategy involves leveraging existing infrastructure, such as process cameras, and deploying models at the edge for real-time decision-making, all while maintaining human oversight. This approach transforms vision AI from isolated pilots into scalable, business-led solutions, fostering trust and driving adoption.

In 2026, RF-DETR emerges as the leading choice for computer vision projects, excelling in object detection and segmentation benchmarks like COCO and RF100-VL, while other models such as SAM 3, DINOv3, GLM-OCR, and Gemini 3.5 Flash excel in their respective tasks. This guide provides a comprehensive comparison of top models across various computer vision tasks including object detection, segmentation, classification, pose estimation, OCR, and visual reasoning, emphasizing the importance of task-specific model selection based on factors such as benchmark performance, real-world applicability, and deployment constraints. Key models like YOLO26, YOLO11, and RF-DETR are highlighted for object detection, while RF-DETR Segmentation, SAM 3, and SAM 2 are recommended for instance segmentation tasks. Vision-language models like Gemini 3.5 Flash and Florence-2 are detailed for tasks requiring visual understanding combined with natural-language processing. The document underscores the importance of evaluating models on real-world data and contextualizing them within specific use cases, encouraging the use of Roboflow's tools for testing, training, and deploying models tailored to custom needs.

Computer vision is an advanced AI field that transforms visual data from images, videos, and live feeds into structured information, allowing software to interpret and act on what it perceives. This technology is utilized across various industries for tasks like defect detection, inventory management, and robotic guidance, using a combination of foundational and custom models. Modern systems, like RF-DETR and DINOv3, work in real-time and adapt to changes in environment and task requirements, offering solutions from object detection to optical character recognition (OCR). Computer vision applications are deployed either on cloud platforms or edge devices, depending on latency, connectivity, and cost considerations, with hybrid systems often combining both for optimal performance. The integration of vision AI in sectors like manufacturing, logistics, agriculture, healthcare, and retail enables real-time decision-making by converting vast amounts of visual data into actionable insights, thus enhancing operational efficiency and decision-making processes.

The tutorial highlights the capabilities of SAM 3, a segmentation model that employs promptable concept segmentation to identify and segment objects in images based on text prompts without needing training or manual input. SAM 3's unique feature is its ability to detect and segment every instance of a concept, such as "helmet" or "safety vest," across an entire scene, which is particularly useful for tasks like construction site PPE inspection. The tutorial details a Roboflow Workflow that utilizes SAM 3 to efficiently generate segmentation masks from text prompts and create AI-powered safety summaries, showcasing the potential of text-prompt segmentation for rapid visual prototyping and pre-labeling datasets. Unlike traditional models that require bounding boxes and training, SAM 3 simplifies the process and provides pixel-accurate masks, making it advantageous for scenarios where objects are small, overlapping, or partially hidden.

Open vocabulary segmentation enables the identification and segmentation of any described object within an image using text prompts without relying on labeled datasets or fixed class lists, allowing for flexible and dynamic object detection. This is exemplified by the SAM 3 model from Meta, which processes text prompts to generate pixel-level masks for various vehicle parts without requiring retraining or a pre-defined set of classes. The method is particularly useful in industries like automated vehicle damage detection, which reached a market size of $1.43 billion in 2025, by replacing manual inspections with computer vision systems that adapt to new parts through simple text descriptions. The tutorial demonstrates how to use SAM 3 within Roboflow Workflows to segment car parts from a text list, outputting annotated images with color-coded masks and detailed coverage reports for each segment. Unlike traditional semantic segmentation, which is limited to fixed class sets, open vocabulary segmentation allows for zero-shot segmentation of arbitrary categories, leveraging models like SAM 3 that integrate vision and language processing. The flexibility of this approach supports various applications beyond automotive, such as segmenting warehouse assets or construction equipment, by merely altering the text prompts without changing the underlying workflow structure.

Video process monitoring is an automated system that utilizes computer vision to extract operational data such as object counts, throughput, speed, and anomalies from camera feeds without human supervision. By integrating RF-DETR for object detection, ByteTrack for tracking with unique IDs, and custom counting logic in Roboflow Workflows, users can establish a real-time analytics pipeline applicable to various fields like manufacturing and logistics. The tutorial demonstrates creating a process monitoring pipeline that processes video feeds, such as highway CCTV footage, to detect and track vehicles, estimate their speed, and count them as they cross a specific line. The setup requires fixed cameras and consistent movement, and it can be adapted for different industrial processes by modifying the model and counting logic. The Roboflow platform allows for deployment on various devices, including edge devices and the cloud, supporting both live and recorded video inputs, and it facilitates scaling by enabling retraining with new data as more observations are gathered.

Computer vision technology is revolutionizing surface defect inspection in manufacturing by utilizing object detection models to quickly identify and classify visual irregularities in materials such as wood, metal, and glass. This automation, exemplified by a tutorial using a custom RF-DETR model in Roboflow Workflows, enables quality teams to efficiently categorize product health and manage defective items on production lines. Traditional manual inspections, reliant on human assessment, can be enhanced by computer vision systems that consistently process images and integrate results into quality-management systems. Applications span various industries, including metal sheet inspection, printed circuit board examination, textile and fabric assessment, glass manufacturing, ceramic tile quality control, and paint and coating defect detection. The tutorial details the creation of a wood surface inspection application, emphasizing the system's ability to detect, count, and categorize defects, visualize results, and log events for future model improvement. By transitioning from mere defect detection to comprehensive visual inspection applications, computer vision facilitates more efficient decision-making regarding product acceptance, review, rework, or rejection, ultimately improving manufacturing processes and product quality.

Product recognition AI is revolutionizing retail, warehousing, logistics, and smart checkout systems by automating the verification of physical products against transaction records like receipts and invoices. By integrating object detection and text recognition technologies, businesses can streamline processes such as checkout validation and order packing, thereby eliminating human errors. Utilizing custom SKU-level detection models and GLM-OCR within Roboflow Workflows enables retailers to ensure planogram compliance, detect out-of-stock items, verify shelf prices, and enhance inventory management. The AI system is not only a detection tool but a comprehensive workflow that supports decision-making and operational accuracy by transforming images from various sources into actionable data. A notable application is the Product-to-Bill Verification System, which identifies products from images, reads receipts via OCR, and verifies product-bill consistency, thus minimizing errors at crucial transaction points. This system can be deployed in various scenarios, including retail checkout, e-commerce packing, warehouse dispatch, and returns, thereby improving efficiency and reducing potential losses.

Computer vision is revolutionizing the inspection of paint defects by utilizing the RF-DETR object detection model to accurately identify and classify surface imperfections such as scratches, bubbling, and orange peel, which are critical for maintaining the quality and protective performance of coatings in industries like automotive and aerospace. By integrating this model into Roboflow Workflows, manufacturers can automate the labeling and counting of defects, streamlining the quality control process. The system, trained on a dataset with 12 types of paint defects, uses preprocessing and augmentation techniques to ensure the model can generalize effectively to new conditions, achieving moderate performance with metrics such as a 67.2% mAP@50. Despite challenges in distinguishing defects with similar visual features, the system offers a scalable solution for consistent inspection, which can be adapted to various manufacturing contexts. Enhancing the model with more representative images and addressing weak-performing classes could further improve its efficacy in specific production environments.

Roboflow has integrated Workflows directly into its Auto Label feature, allowing users to run custom, multi-model pipelines serverlessly for annotating unannotated images. This integration provides the flexibility to utilize various AI models and tools within the annotation interface, facilitating the creation of multi-step pipelines, model ensembles, and advanced consensus rules without being restricted to a single model. By employing a rules-based consensus engine, users can ensure more accurate labeling as different models, such as Google Gemini, OpenAI's GPT, and Anthropic's Claude, work together to eliminate uncorrelated errors. This process enhances data labeling by providing greater control over data perception, filtering, and categorization, ultimately reducing human error and scaling up dataset sizes efficiently. The ability to construct advanced guardrails through this integration marks a significant advancement in data engineering pipelines, offering a customizable and scalable solution for teams aiming to optimize their data annotation processes.

Open-source computer vision models offer developers powerful tools for tasks like object detection, segmentation, depth estimation, and more, without relying solely on closed APIs. The top models in 2026 include RF-DETR, YOLO26, SAM 3, Grounding DINO, GLM-OCR, Depth Anything 3, CLIP, Florence-2, and Qwen3.5, each excelling in specific areas such as real-time detection, text extraction, and multimodal understanding. Licensing varies, with some models under permissive licenses like Apache 2.0 and MIT, while others have custom licenses requiring careful review. Roboflow supports these models by providing a unified platform for dataset management, annotation, training, and deployment, allowing developers to test and evaluate models on their data to find the best fit for accuracy, speed, cost, and deployment needs. Practical considerations include dataset quality, prompt design, and licensing, with the recommendation to test multiple models to balance accuracy, speed, and application requirements effectively.

OpenAI has launched the GPT-5.6 model family, consisting of three distinct tiers: Sol, Terra, and Luna, each catering to different needs in terms of cost, speed, and accuracy. Sol is designed for maximum reasoning and deep contextual understanding, excelling in complex tasks with high accuracy, while Luna focuses on high-volume, real-time tasks with lower cost and latency, and Terra provides a balanced middle ground for everyday work. These models are now available for testing in the Roboflow Playground, where users can compare them against other models using their own images and prompts. The models also support Programmatic Tool Calling, enhancing efficiency by executing sequences of tool calls without unnecessary data routing. The Playground offers a user-friendly environment to evaluate the performance of these models on custom data, highlighting their real-world capabilities and economic trade-offs.

Vision Events serves as a comprehensive hub for managing and improving computer vision systems by centralizing all the predictions, images, and metadata produced by these systems. It facilitates collaboration between vision systems and experts such as quality engineers and maintenance planners by ensuring data is accurate, accessible, and actionable. Key features include closing the feedback loop through operator feedback, built-in dashboards for monitoring deployments, and plain-language querying via an API and the Roboflow MCP Server. This allows teams to continuously improve their models by integrating real-world feedback into training datasets, ensuring that systems adapt to changes such as environmental shifts or new processes. Vision Events thus enhances the utility of vision systems by making predictions both trustworthy and usable for decision-making and operational improvements.

Pose estimation is a vital computer vision task that predicts the location of keypoints on an object, such as a person's joints, and connects them into a skeleton to interpret body posture or movement. This guide explores the function and application of pose estimation, emphasizing its life-saving potential in distinguishing between swimming and drowning by analyzing body posture rather than mere presence. The guide details the training of an RF-DETR Keypoint model on a dataset of underwater images to create a workflow capable of returning skeleton overlays and alerts based on a person's pose. It highlights the model's ability to differentiate between a swimmer and a person in distress, underscoring the importance of pose estimation in safety systems, sports, physical therapy, and industrial applications. The guide also discusses the architecture of pose models, such as top-down and bottom-up approaches, and emphasizes the advantages of modern transformer-based models for real-time inference. Additionally, it outlines a practical workflow using Roboflow for training and deployment, enabling continuous monitoring and improvement of the model's accuracy over time.

Brian Ton, senior laboratory manager at Florida Crystals, shares his insights on implementing visual AI in complex manufacturing, emphasizing the importance of change management over technology choice. He highlights the necessity of a robust validation and verification system, ensuring accessibility across organizational levels, and maintaining proximity to subject matter experts. Ton underscores the significance of an operator feedback loop in AI systems, which acts as quality control for AI tools. His approach involves achieving small victories to demonstrate value and build trust in visual AI systems, rather than pursuing overly ambitious projects that can lead to discouragement and abandonment. Persistence, rather than technology itself, is identified as a key factor in successfully integrating visual AI into standard operations.

Dimensional defects, which occur when a part's size, spacing, or alignment deviates from its intended specifications, can significantly impact product functionality and quality, despite being invisible to the naked eye. The article discusses how to automate the detection of these defects using a Roboflow Workflow that employs computer vision and a Python script to measure and evaluate the dimensions of parts, such as the spacing between mounting holes on a steel bracket. By using Google Gemini to detect features and a custom Python block to compare actual measurements against tolerances, the workflow provides a pass or fail verdict. This automation enables manufacturers to perform consistent and scalable inspections, reducing reliance on manual quality control and allowing for adaptation to various dimensional inspection tasks beyond hole spacing, such as checking gaps, component alignment, and overall dimensions.

A "detect anything" model, such as Segment Anything Model 3 (SAM3), revolutionizes the traditional object detection workflow by allowing users to locate and identify objects in images based on plain-language prompts rather than relying on a fixed, pre-trained set of classes. This approach is particularly beneficial for environments like warehouses, where workers can significantly reduce non-productive time spent searching for assets. In this tutorial, SAM3 is used within Roboflow Workflows to create a dynamic asset search system that responds to text prompts like "forklift" or "pallet," instantly providing bounding boxes, labels, and counts of detected items without needing labeled data or retraining. The system's versatility means that adding a new object requires only a change in the text prompt, offering immediate results and making it ideal for prototyping or situations with rare or variable objects. While this model offers flexibility and rapid deployment, it trades off some accuracy compared to traditional models, which may be more precise for specific, consistently labeled classes.

The tutorial explores the implementation of predictive maintenance using vision AI by integrating RF-DETR and Gemini 2.5 Pro within Roboflow to detect equipment defects before failure occurs. This approach leverages cameras and trained computer vision models to identify early-warning signs of equipment issues such as surface defects and corrosion, preventing costly unplanned downtime. The process involves training an RF-DETR model with annotated bearing images to classify conditions like good, scratch, rust, or grease, and deploying it in Roboflow Workflows to automate visual inspections. The model's predictions are analyzed by Gemini 2.5 Pro, which provides maintenance insights, including the likely causes and severity of detected conditions, and recommends actions. This integration of computer vision with maintenance reasoning enables efficient detection and assessment of equipment health, offering a scalable solution that can be extended to various industrial applications beyond bearings.

Alexei Alexandrovich introduces a tool designed to democratize the technology behind soccer's offside verdicts by making it accessible for general use. Unlike the expensive semi-automated systems used in professional games, which rely on a rig of synchronized cameras and a control team, this new system can analyze any video footage to determine offside positions using a single camera. The process involves uploading a clip, detecting pitch lines, identifying relevant players, and reconstructing them in 3D using tools such as RF-DETR and SAM3D Body, which work together to provide a 3D representation of the scene. Although not as precise as FIFA's multi-camera VAR, the system offers promising results, running on a phone video for minimal GPU costs. It is available for public use through a Hugging Face space, potentially bringing advanced VAR technology to local games and amateur footage.

Claude Sonnet 5, released by Anthropic on June 30, 2026, is a mid-tier model in the Claude 5 family that provides a lateral move in computer vision capabilities compared to its predecessor, Sonnet 4.6, showing no improvement in benchmark accuracy. It achieves a 70% pass rate on Roboflow Vision Evals, matching Sonnet 4.6, yet it falls short in document understanding and object counting, with scores of 67% and 20% respectively, and is outperformed by competitors like Gemini 3.5 Flash, which leads with a 79% vision score. Despite introductory pricing through August 31, 2026, Sonnet 5 is not the most cost-effective option for vision accuracy per dollar, as it is outstripped by models like Gemini 3.5 Flash that offer superior performance at a lower cost. While Sonnet 5 is suitable for visual question answering and general image understanding, it is not the top choice for tasks like detection, counting, and segmentation, where a fine-tuned RF-DETR model is more accurate and cost-efficient.

Artificial intelligence (AI) in 2026 is categorized into systems that generate and reason with information, like language models, and those that perceive and act on the physical world, such as computer vision and robotics, with many applications combining these capabilities. AI development relies heavily on machine learning, where models learn from data rather than manually programmed rules, making them effective for specialized tasks like detecting defects or predicting equipment failures. While current AI systems are considered narrow in scope, focusing on specific tasks, the concept of Artificial General Intelligence (AGI) remains hypothetical, sparking ongoing debate among researchers. The development of modern AI began long before computers, with milestones like IBM's Deep Blue and Watson marking significant advancements, and contemporary AI increasingly focuses on smaller and more efficient models deployed across various systems. AI's real-world applications span manufacturing, healthcare, agriculture, logistics, and retail, using generative and vision AI to transform data into tangible outcomes, while challenges like data privacy and governance are addressed through regulations like the EU AI Act and frameworks like the NIST AI Risk Management Framework.

Natural language image annotation enhances the efficiency of labeling images by allowing users to describe objects in plain text rather than manually drawing boxes, significantly speeding up the data preparation phase in computer vision projects. This tutorial introduces two approaches: using Roboflow's Auto Label feature with Meta's Segment Anything Model (SAM3) to generate pixel-level segmentation masks directly within the UI without coding, and employing Autodistill with Grounding DINO to automate local image labeling using text prompts in Python. Both methods culminate in training an RF-DETR model, demonstrating that natural language annotation can efficiently transform raw images into a labeled dataset with minimal manual intervention. While Roboflow's Auto Label is ideal for projects requiring detailed segmentation masks, Autodistill suits those preferring a code-based, local solution, allowing users to streamline the labeling process and focus manual efforts on complex or difficult-to-describe classes.

Roboflow CEO Joseph Nelson discusses the challenges and strategies of deploying computer vision systems in industrial settings, emphasizing the importance of a customized model that improves through active learning and is integrated with business systems for real-time decision-making. The deployment process is broken into three critical parts: obtaining relevant data and footage, developing a model tailored to specific business needs, and ensuring the model's outputs are actionable within the enterprise's operational systems. Nelson highlights the necessity for models to adapt to real-world conditions, with active learning and real-time transformer models now making such deployments more feasible. The conversation also covers strategic approaches such as the barbell strategy for executive buy-in and the center-of-excellence model, which helps enterprises scale up their AI initiatives effectively. Roboflow supports enterprises in transitioning from pilot projects to full-scale production, leveraging the current market dynamics where Vision AI is becoming increasingly mainstream, though the opportunity for differentiation is diminishing.

Manufacturing vision systems integrate cameras, lighting, and software to automate decision-making on production lines, such as quality checks, robotic guidance, and machine monitoring, transforming what was once a human task into a machine-driven process. These systems range from simple rule-based smart cameras for stable checks to sophisticated AI-driven models that identify evolving defects without requiring frequent reconfiguration. The technology is prevalent across modern manufacturing floors, enabling consistent inspection accuracy and freeing human workers for more complex tasks. The systems are adaptable, with options spanning from smart cameras to PC-based configurations, each suited to different manufacturing needs, with the potential to significantly reduce costs associated with defects and rework. Integration with existing plant systems like PLCs and MES is crucial for turning these vision systems from standalone tools into operational assets, and platforms like Roboflow facilitate the deployment of custom-trained AI models without necessitating an in-house machine learning team.

Object detection models have become crucial in various real-time applications such as manufacturing, traffic systems, and edge AI due to their ability to identify defects, monitor inventory, and count objects. The choice of model largely depends on balancing speed, accuracy, and deployment needs. In 2026, transformer-based models like RF-DETR are highlighted for their high accuracy and minimal post-processing latency, making them suitable for diverse datasets and environments. Other models like YOLO26, Roboflow 3.0, YOLOv12, and RT-DETR emphasize improvements in speed and accuracy without heavy post-processing, making them ideal for low-latency deployments. Roboflow's ecosystem supports various models, offering tools for training and deploying them on custom datasets. The guide emphasizes using Roboflow Workflows to compare model performance on specific hardware and datasets, enabling users to select the most efficient model for their needs. Profiling tools within Roboflow allow for detailed timing analysis, helping optimize model deployment on different hardware setups.

A two-stage detector is an object detection model that processes images in two steps: identifying potential object regions and then classifying and refining these regions, enhancing accuracy and reliability in complex scenes with small, overlapping, or indistinct objects. This approach, used in the R-CNN family of models, includes advancements such as Fast R-CNN, which shares feature maps to improve speed, Faster R-CNN, which incorporates a Region Proposal Network, and Mask R-CNN, which extends functionality to instance segmentation. Despite their higher accuracy, two-stage detectors are slower than one-stage detectors, which predict bounding boxes and class labels in a single pass. For practical applications, tools like Roboflow assist in dataset preparation through image annotation, versioning, and exportation, which can then be used to train models like Faster R-CNN in environments such as Colab.

Computer vision is revolutionizing the manufacturing industry by automating the defect inspection process, thereby enhancing speed and consistency in quality control. By integrating object detection models like Roboflow's RF-DETR with vision-language models such as Gemini, manufacturers can create custom pipelines that detect, classify, and report product flaws in real time. This automation is crucial as manual inspection becomes challenging with increased production speeds. The defect inspection process aims to identify flaws like cracks, incorrect dimensions, or cosmetic imperfections before products reach consumers, thus minimizing waste, production delays, and customer dissatisfaction. Different types of manufacturing defects—surface, dimensional, assembly, and cosmetic—require distinct computer vision models, including object detection, segmentation, and anomaly detection. These models can be part of a broader inspection workflow that not only identifies defects but also provides natural language explanations and integrates with factory equipment for real-time action. Whether deployed in the cloud, at the edge, or offline, these systems are adaptable to a wide range of manufacturing applications, from electronics to pharmaceuticals, illustrating the transformative impact of computer vision in modern manufacturing.

Cost of poor quality (COPQ) measures the financial losses caused by defects, encompassing internal failures such as scrap, rework, re-inspection, and quality-related downtime, as well as external failures including warranty claims, returns, recalls, replacement costs, and lost customers. Unlike prevention and appraisal costs, COPQ focuses on failure costs and can be calculated from records in ERP systems, stoppage logs, and finance or claims data; an example $80 million manufacturing plant recorded $4.24 million in annual failure costs, plus $420,000 in manual inspection labor. The article argues that defects become substantially more expensive the later they are found, following the approximate 1-10-100 rule from station-level detection to end-of-line detection to customer discovery. It presents camera-based computer vision as a means to shift from sample-based to continuous inspection, catch defects earlier, reduce customer escapes, and redirect inspectors toward disposition and root-cause work. Roboflow promotes its image-trained, deployable computer vision platform and active-learning capabilities as tools for implementing such systems, while recommending that quality leaders use a conservative, finance-supported COPQ estimate and begin with a limited pilot focused on one production line and defect class.