May 2023 Summaries

The May 25th Computer Vision Meetup featured presentations from Opendoor, Deci AI, and Wildlife.ai. Topics covered included the application of computer vision in real estate at Opendoor, the release of YOLO-NAS by Deci AI, and the use of artificial intelligence and computer vision for wildlife conservation by Wildlife.ai. The event also highlighted the growth of the Computer Vision Meetup network to over 4,000 members and announced upcoming events in various locations.

This week's FiftyOne tips and tricks blog covers various topics such as computing image embeddings using a PyTorch model, exporting datasets with absolute paths, getting data into the FiftyOne App, downloading annotated videos, deleting samples from a dataset, and more. The blog also highlights the growing community of FiftyOne users on Slack, GitHub, Stack Overflow, and Reddit.

The annual IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) is set to take place in Vancouver, Canada from June 18th – June 22nd. With over 2300 accepted papers, the conference will feature a range of topics including human-centric visual tasks, text-to-image diffusion models, neural radiance field training, open-set grounded text-to-image generation, and more. The top 10 papers to look out for include "Beyond Appearance," "DreamBooth," "F2-NeRF," "GLIGEN," "ImageBind," "Mask DINO," "MobileNeRF," "Planning-oriented Autonomous Driving," "SadTalker," and "VideoFusion." These papers represent significant advancements in the field of computer vision, offering innovative solutions to various challenges.

This week's tips and tricks blog covers various aspects of using FiftyOne, an open source machine learning toolset for computer vision models. Topics discussed include deleting specific labels in a dataset, improving the performance of extracting and downloading object patches, creating dataset views after filtering, loading large thumbnails into the FiftyOne App quickly, adding custom sidebar groups, and joining the FiftyOne community. The blog also provides useful links to the FiftyOne Docs for more detailed information on these topics.

The annual IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) is a significant event in the field of computer vision. With thousands of researchers and engineers participating, it showcases pioneering projects and papers. This year's conference will take place from June 18th – June 22nd in Vancouver, Canada, featuring 2359 accepted papers, 100 workshops, 33 tutorials, and Flash Sessions. The data gathered from these papers provides insights into the current state of computer vision, its trends, and future directions.

The May 11, 2023 Computer Vision Meetup featured presentations on topics such as visualizing defects in Amazon's ARMBench dataset using embeddings and OpenAI's CLIP model, the role of symmetry in human and computer vision, and machine learning for fast, motion-robust MRI. The event also included a Q&A session with speakers Sven Dickinson and Nalini Singh. The next Computer Vision Meetup is scheduled for May 25 at 10 AM IST, featuring talks on YOLO-NAS, Wildlife Watcher, and applying computer vision to real estate at Opendoor. Membership in the Computer Vision Meetup has grown to over 4,000 members across 13 locations worldwide.

This week's FiftyOne tips and tricks blog covers various topics such as counting classes using the FiftyOne API, adding predictions to videos, deleting samples with uniqueness less than a certain value, adding VOC labels to samples, getting started with model predictions in FiftyOne, and joining the FiftyOne community. The blog provides code snippets and explanations for each topic, making it easier for users to understand and implement these techniques in their work.

The FiftyOne community has been thriving with various organizations using it for their computer vision projects. Allstate India uses FiftyOne to segment vehicle parts, detect damages, and generate reports for adjusters. Secury360 leverages FiftyOne to manage large image and video datasets and improve its surveillance model. The community is encouraged to share their success stories with FiftyOne and claim rewards. In April, the release of FiftyOne 0.20.1 brought over 90 enhancements and fixes. Community members have also contributed significantly to the project's development.

Deci AI has released a new state of the art object detection model named YOLO-NAS. This model achieves higher mean average precision than prior models running with the same latency. In this tutorial, we will show how to generate predictions with YOLO-NAS and load them into FiftyOne, an open source computer vision toolset for curating better data and building better models. The tutorial includes setup instructions, loading a dataset from the FiftyOne Dataset Zoo, generating YOLO-NAS predictions, converting bounding boxes, adding YOLO-NAS detection predictions to the dataset, evaluating the quality of object detection predictions against the "ground truth", and more.

In this blog post, we explored Amazon's recently released computer vision dataset for training "pick and place" robots using the open-source FiftyOne toolset. The ARMBench dataset is the largest computer vision dataset ever captured in an industrial product-sorting setting, featuring over 235,000 pick and place activities on 190,000 objects. We focused on the Image Defect Detection subset of data and utilized FiftyOne to visualize it. Additionally, we created embeddings with the OpenAI CLIP model to explore defects further. The practical applications of "pick and place" robots include manufacturing, packaging, sorting, and inspection tasks that require speed and accuracy.

The article discusses the integration of Weights & Biases (W&B) and FiftyOne for model and dataset tracking, visualization, and evaluation. It highlights how these tools can be used to co-develop high-quality datasets and models by iteratively improving them together. The author demonstrates this process using a food detection task with Hugging Face's DETR model. They cover the steps of curating a custom dataset with FiftyOne, integrating W&B and FiftyOne in the model training loop, visualizing model predictions in FiftyOne, sweeping over hyperparameters with W&B, and finding the best model with both tools. The integration between W&B and FiftyOne allows users to view high-level model evaluation results in W&B and explore corresponding model predictions and evaluations on actual samples in FiftyOne.