Research map / Video / Video segmentation and tracking
CamSAM2: Segment Anything Accurately in Camouflaged Videos
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CamSAM2, which adds a learnable decamouflaged token and IOF, EOF, and OPG modules to a frozen SAM2 to improve video camouflaged object segmentation with point, box, and mask prompts.
Problem
SAM2 segments camouflaged objects in videos poorly, especially with simple point or box prompts, because its features are optimized for natural scenes and its memory design keeps only coarse features.
Approach
Keep all SAM2 weights frozen and add a learnable decamouflaged token to the mask decoder, plus fusion of high-resolution early-layer features (IOF), cross-attention to object prototypes generated by FPS and k-means from previous frames (EOF/OPG), and average the SAM2 and CamSAM2 mask logits.
Key results
- On MoCA-Mask with Hiera-T and 1-click prompt, CamSAM2 reaches 64.3 mDice vs. 52.1 for SAM2 (+12.2).
- On SUN-SEG-Hard with mask prompt, mDice rises from 61.0 (SAM2) to 80.6 (CamSAM2), a +19.6 gain.
- CamSAM2 adds about 0.5M parameters and about 6 ms per frame latency (7.2%) over SAM2 on Hiera-T.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CamSAM2's decamouflaged token, IOF, EOF, and OPG | SAM2, Hiera-T · MoCA-Mask, 1-click | mDice: 52.1 -> 64.3 (+12.2) | Helped |
| Add CamSAM2 | SAM2, Hiera-T · CAD, 1-click | mDice: 59.2 -> 62.6 (+3.4) | Helped |
| Add CamSAM2 | SAM2, Hiera-T · SUN-SEG-Easy, mask | mDice: 73.6 -> 84.3 | Helped |
| Add CamSAM2 | SAM2, Hiera-T · SUN-SEG-Hard, mask | mDice: 61.0 -> 80.6 | Helped |
| Add the decamouflaged token (ablation) | SAM2, Hiera-T · MoCA-Mask, 1-click | mDice: 52.1 -> 54.9 | Helped |
| Add IOF to the decamouflaged-token variant (ablation) | CamSAM2 · MoCA-Mask, 1-click | mDice: 54.9 -> 55.2 | Helped |
| Add EOF to the decamouflaged-token + IOF variant (ablation) | CamSAM2 · MoCA-Mask, 1-click | mDice: 55.2 -> 55.9 | Helped |
| Add OPG to the decamouflaged-token + IOF + EOF variant (ablation) | CamSAM2 · MoCA-Mask, 1-click | mDice: 55.9 -> 64.3 | Helped |
| Use cosine rather than Euclidean distance for OPG k-means (ablation) | CamSAM2 · MoCA-Mask, 1-click | mDice: 61.9 -> 64.3 | Helped |
| Set OPG prototype count to 5 rather than 3 or 7 (ablation) | CamSAM2 · MoCA-Mask, 1-click | mDice: 60.2 (k=3); 64.3 (k=5); 60.6 (k=7) | Helped |
| Use 3 rather than 1 decamouflaged token (ablation) | CamSAM2 · MoCA-Mask, 1-click | mIoU: 54.6 -> 53.3 | Hurt |
| Use FPS rather than Average Sampling for OPG centers, with k-means clustering (ablation) | CamSAM2 · MoCA-Mask, 1-click | mIoU: 52.1 -> 54.6 | Helped |
Surprising findings
- Averaging SAM2 token and decamouflaged token logits improves the SAM2 token's own output (SAM2 token mIoU 53.3 vs. 44.8 for the original SAM2), even though its parameters are unchanged.
Limitations
- Does not address SAM2's challenges with shot changes and long occlusions.
- Inherits SAM2's architecture and known limitations.
- Evaluation uses ground-truth-derived prompts on the first frame, which may raise fairness concerns.
Experimental setup
Datasets: MoCA-Mask MoCA-Mask-Pseudo SUN-SEG CAD
Benchmarks: MoCA-Mask CAD SUN-SEG-Easy SUN-SEG-Hard
Metrics: mDice mIoU S-measure weighted F-measure MAE E-measure F-measure
Baselines: SAM2 SAM2-FT SAM-PM ZoomNeXt SLT-Net-LT SAM2-Adapter EGNet SINet-V2
Builds on: SAM2 SAM HQ-SAM SAM2-Adapter
Closest papers by meaning
- Endow SAM with Keen Eyes: Temporal-spatial Prompt Learning for Video Camouflaged Object Detection (CVPR 2024 · similarity 0.88)
- Camouflage Anything: Learning to Hide using Controlled Out-painting and Representation Engineering (CVPR 2025 · similarity 0.87)
- SAM-I2V: Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost (CVPR 2025 · similarity 0.87)
- SAM 2: Segment Anything in Images and Videos (ICLR 2025 · similarity 0.86)
- SAMWISE: Infusing Wisdom in SAM2 for Text-Driven Video Segmentation (CVPR 2025 · similarity 0.86)
- Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.