Research map / Detection and segmentation / Tracking
Efficient Track Anything
Yunyang Xiong, Chong Zhou, Xiaoyu Xiang, Lemeng Wu, Chenchen Zhu, Zechun Liu, Saksham Suri, Balakrishnan Varadarajan, Ramya Akula, Forrest Iandola, Raghuraman Krishnamoorthi, Bilge Soran and 1 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace hierarchical Hiera image encoder with plain ViT-Small | SAM 2 architecture · SA-V test | J&F: 74.7 -> 74.5; FPS 43.8 -> 85.0 | No clear effect |
| add average-pooled efficient memory cross-attention (2x2 window) | EfficientTAM-S · SA-V test / iPhone 15 | J&F and latency: J&F 74.5 -> 74.0; latency 1010.8 -> 450 ms | Mixed |
| add object pointer tokens to cross-attention (ablation) | EfficientTAM · SA-V test | J&F: 72.1 -> 74.5 | Improved |
| pool only spatial memory tokens rather than all memory tokens (ablation) | efficient cross-attention · SA-V test | J&F: 71.7 -> 74.0 | Improved |
| use 4x4 pooling window instead of 2x2 (ablation) | efficient cross-attention · SA-V test | J&F: about 1 J&F drop with marginal speed gain | Hurt |
| use linear attention (associative property) instead of pooled cross-attention (ablation) | efficient cross-attention · SA-V test | J&F: >10 J&F drop | Hurt |
| use average pooling with log-window constant instead of Linformer-style variant without constant (ablation) | efficient cross-attention · SA-V test | J&F: 73.4 -> 74 | Improved |
| use average pooling instead of local windowed cross-attention (ablation) | efficient cross-attention · SA-V test | J&F: 72.4 -> 74.0 | Improved |
| reduce input resolution from 1024x1024 to 512x512 (ablation) | EfficientTAM-S · SA-V test | J&F: 74.5 -> 71.5; iPhone latency 1010.8 -> 80.6 ms | Hurt |
Closest papers by meaning
- Tracking Every Thing in the Wild (ECCV 2022 · similarity 0.85)
- [Re] Transparent Object Tracking Benchmark (NeurIPS 2022 · similarity 0.80)
- ARTrackV2: Prompting Autoregressive Tracker Where to Look and How to Describe (CVPR 2024 · similarity 0.73)
- TGFormer: Transformer with Track Query Group for Multi-Object Tracking (AAAI 2025 · similarity 0.73)
- Focusing on Tracks for Online Multi-Object Tracking (CVPR 2025 · similarity 0.73)
- RTracker: Recoverable Tracking via PN Tree Structured Memory (CVPR 2024 · similarity 0.72)
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.
