Скачать или смотреть SAM 3: Open-Vocabulary Object Segmentation

In this AI Research Roundup episode, Alex discusses the paper: 'SAM 3: Segment Anything with Concepts' SAM 3 is a unified Promptable Concept Segmentation model that can detect, segment, and track all instances of arbitrary visual concepts using short noun phrases, image exemplars, or both. It tackles open-vocabulary instance segmentation and tracking in images and short videos, returning masks and unique identities for every object matching a concept prompt, even under ambiguity and negative examples. The model uses a shared perception encoder for both a DETR-style detector and a memory-based video tracker, along with a presence token that separates recognizing whether a concept is present from localizing where it is. SAM 3 also supports interactive refinement with exemplar prompts and clicks, and can be integrated with multimodal language models for more complex visual-language queries. Paper URL: https://arxiv.org/abs/2511.16719 #AI #MachineLearning #DeepLearning #ComputerVision #InstanceSegmentation #VideoTracking #OpenVocabulary

Resources:
GitHub: https://github.com/facebookresearch/sam3