February 2026 Summaries

Click-to-Segment is a feature integrated into FiftyOne Labs that revolutionizes segmentation annotation by utilizing AI-powered interactive segmentation tools. This feature allows users to create segmentation masks with just a few clicks, leveraging promptable segmentation models like the Segment Anything Model (SAM) from the FiftyOne Model Zoo. Users can enable Click Capture to generate instance-level keypoints for fine-grained control over segmentation masks, which are then displayed within the FiftyOne App. The feature supports both positive and negative point prompts, offering flexibility in mask creation, and includes a Segment with Prompts operator for batched inference. This tool addresses the limitations of traditional polygon-based annotation methods, making it easier to build high-quality segmentation datasets directly within the existing data management workflow in FiftyOne.

FiftyOne, an open-source ML platform by Voxel51, addresses inefficiencies in data annotation workflows for visual AI by integrating strategic data selection, automation, and human expertise into a unified system that enhances model performance and reduces costs. Traditional approaches that prioritize labeling vast amounts of data often result in wasted budgets and persistent errors, as they miss rare scenarios crucial for robust models. FiftyOne's new capabilities include manual annotation for 2D and 3D labels directly within the platform, eliminating tool-switching delays, and leveraging techniques like zero-shot coreset selection to prioritize data samples that most improve model performance. By unifying data curation, annotation, and model evaluation, FiftyOne enables ML teams to systematically improve model accuracy while maintaining contextual integrity throughout the development process.

The Few-Shot Learning plugin in FiftyOne Labs is designed to enhance image retrieval and identify model failures or niche cases within partially labeled or unlabeled datasets by leveraging a binary notion of similarity, where users label a small set of positive and negative samples. The plugin employs embeddings from models like ResNet, CLIP, and DINOv2, utilizing a prototype-style scoring system to rank the rest of the dataset. This lightweight workflow allows for rapid iteration by sampling and stabilizing working subsets, making it particularly useful for mining edge cases and growing niche classes. It simplifies the process by training a Rocchio-inspired prototype classifier, which writes back predictions to the dataset, enabling users to refine and improve their models iteratively within the FiftyOne App's interactive environment.

LightOnOCR-2-1B is a groundbreaking 1-billion-parameter vision-language model that transforms document OCR by offering state-of-the-art performance while being significantly more compact than previous models. Traditional OCR systems have faced challenges with complex layouts, multilingual content, and scientific notation, but LightOnOCR-2 excels by using an end-to-end architecture that eliminates the need for multi-stage processing and complex post-processing. This model achieves high accuracy with a smaller size, processing documents at 5.71 pages per second on H100 GPUs, and is particularly adept at handling European languages, especially French, due to its advanced training techniques. It integrates seamlessly with FiftyOne for efficient batch processing of document datasets, making it highly practical for real-world applications. Released under the Apache 2.0 license, LightOnOCR-2 is accessible for both research and commercial use, offering new opportunities for document digitization across various industries.

Document understanding is a complex challenge in computer vision, traditionally relying on OCR systems that excel at character recognition but struggle with complex document structures. GLM-OCR, a multimodal AI model, offers a significant advancement by integrating vision and language understanding to semantically process documents, preserving their structure in formats like Markdown, JSON, or LaTeX. This approach enables efficient parsing of tables, formulas, and layouts, which traditional OCR systems cannot handle without extensive post-processing. The integration of GLM-OCR with FiftyOne enhances its capabilities through efficient batching, dataset management, and visualization, making it suitable for diverse applications such as financial document processing, medical record digitization, and legal document analysis. The system's lightweight design allows deployment on consumer hardware, and its open-source nature facilitates easy integration into existing workflows. By transitioning from character recognition to structure-first extraction, GLM-OCR represents a paradigm shift in building document processing pipelines, offering robust solutions for extracting structured data directly for downstream applications.

In the NVIDIA AI Podcast, experts discuss the challenges facing autonomous vehicle (AV) development, emphasizing that simply accumulating more data is not sufficient for progress. Instead, the focus should be on using smarter data, particularly in areas where models show weaknesses. Many AV companies struggle with data integrity due to issues like misaligned cameras and uncalibrated sensors, which can degrade model performance. Advances in neural reconstruction have made old drive logs valuable, allowing teams to create safety-critical scenarios without relying on real-world conditions. The podcast highlights the importance of integrated workflows over siloed teams to enhance AV development and mentions tools like NVIDIA Omniverse and NVIDIA Cosmos that aid in generating realistic simulations. Voxel51's Physical AI Workbench is presented as a solution to transform raw sensor data into simulation-ready datasets, addressing these challenges effectively.

C-RADIOv4 is an advanced vision foundation model developed by NVIDIA Labs, which integrates the capabilities of multiple state-of-the-art models—SigLIP2, DINOv3, and SAM3—into a single architecture through innovative multi-teacher distillation. This unified model excels in tasks such as zero-shot classification, dense perception, and segmentation, offering competitive performance with significantly fewer parameters than its predecessors. Key technical innovations include stochastic resolution training, shift equivariance, and ViTDet mode for efficient high-resolution processing. C-RADIOv4 is designed to work seamlessly with FiftyOne, a platform that facilitates computer vision workflows by providing tools for dataset exploration, visualization, similarity search, and more. Its application spans various domains, including autonomous vehicles, robotics, and document processing, making it a versatile solution for both research and commercial purposes.