SAM3
SAM3 (Segment Anything Model 3) is a visual segmentation foundation model developed and open-sourced by Meta Research, representing an iteration of the SAM model series. The model supports image segmentation driven by multiple prompt types, including points, bounding boxes, masks, and natural language. It possesses zero-shot generalization capability, enabling precise segmentation of arbitrary objects in images without task-specific fine-tuning, and is applicable to medical imaging, autonomous driving, video understanding, and other visual perception scenarios.
Core Facts
Timeline (last 90 days)
FastBench is constructed using a trajectory-driven data pipeline, comprising four steps: generating QA from high-frame-rate clips, filtering out questions answerable at low frame rate, using SAM3 and CoTracker3 for trajectory answer verification, and three rounds of manual inspection.
SAM3 is a new-generation version of Meta's Segment Anything model series, supporting open-vocabulary segmentation and video object tracking based on text prompts.
Using image masks did not improve reconstruction quality—artifacts remained or were even more pronounced—but the final file size was reduced by approximately 50%.
All Facts (6)
SAM3 is the third-generation product of the SAM series, with its core breakthrough being the extension of segmentation capability from static images to dynamic video and the introduction of natural language understanding
65%UnverifiedSAM3的零样本能力基于视觉-语言对齐训练，通过语义嵌入空间构建实现对未见过类别的泛化
90%UnverifiedSAM3采用基于Transformer的时序建模架构，通过跨帧注意力机制（Cross-frame Attention）维持目标的语义一致性
85%UnverifiedFastBench is constructed using a trajectory-driven data pipeline, comprising four steps: generating QA from high-frame-rate clips, filtering out questions answerable at low frame rate, using SAM3 and CoTracker3 for trajectory answer verification, and three rounds of manual inspection.
50%UnverifiedSAM3 is a new-generation version of Meta's Segment Anything model series, supporting open-vocabulary segmentation and video object tracking based on text prompts.
50%UnverifiedUsing image masks did not improve reconstruction quality—artifacts remained or were even more pronounced—but the final file size was reduced by approximately 50%.
50%