Research map / Architectures and efficiency / Attention and transformers
Exploiting Lightweight Hierarchical ViT and Dynamic Framework for Efficient Visual Tracking
Ben Kang, Xin Chen, Jie Wen Zhao, Chunjuan Bo, Dong Wang, Huchuan Lu
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 |
|---|---|---|---|
| fuse Max, Mid and Min features with the Bridge Module instead of using Max only (ablation) | HiT-Base · LaSOT (ablation, 500 epochs) | AUC: 62.1 -> 63.7 | Improved |
| add the deep Min feature to Max+Mid fusion (ablation) | HiT-Base · LaSOT (ablation, 500 epochs) | AUC: 62.6 -> 63.7 | Improved |
| add the shallow Max feature to Mid+Min fusion (ablation) | HiT-Base · LaSOT (ablation, 500 epochs) | AUC: 60.3 -> 63.7 | Improved |
| add the Mid feature to Max+Min fusion (ablation) | HiT-Base · LaSOT (ablation, 500 epochs) | AUC: 58.8 -> 63.7 | Improved |
| use Shrink Attention for downsampling instead of a stride-2 convolution (ablation) | HiT-Base · LaSOT / TrackingNet / GOT-10k | AUC: decreases of 5.6%, 2.8%, and 6.2% on LaSOT, TrackingNet, and GOT-10k | Improved |
| apply training-free DyTracker acceleration (DyHiT-Route1 plus router) to OSTrack-256 | OSTrack-256 · LaSOT | speed and AUC: 70 fps -> 110 fps, 1.57x; AUC 69.1 -> 69.5 | Improved |
| apply training-free DyTracker acceleration to SeqTrack-B256 | SeqTrack-B256 · LaSOT | speed and AUC: 2.68x speedup; AUC 69.9 -> 69.9 | Improved |
| apply training-free DyTracker acceleration to TransT | TransT · TrackingNet / LaSOT / GOT-10k | averaged AUC change and speed: -0.50 average change, 1.86x speed | Mixed |
| apply training-free DyTracker acceleration to STARK-ST101 | STARK-ST101 · TrackingNet / LaSOT / GOT-10k | averaged AUC change and speed: +0.03 average change, 2.93x speed | Improved |
| apply training-free DyTracker acceleration to OSTrack-256 | OSTrack-256 · UAV123 | AUC: 68.3 -> 70.8 | Improved |
| dynamic routing with router threshold 0.75 instead of full model only (threshold 1) (ablation) | DyHiT · VOT2021 real-time | EAO: 0.252 -> 0.253 | No clear effect |
| use only Route1 (threshold 0) instead of full model (ablation) | DyHiT · VOT2021 real-time | EAO: 0.252 -> 0.250 | Hurt |
Closest papers by meaning
- Exploring Lightweight Hierarchical Vision Transformers for Efficient Visual Tracking (ICCV 2023 · similarity 0.94)
- Two-stream Beats One-stream: Asymmetric Siamese Network for Efficient Visual Tracking (AAAI 2025 · similarity 0.85)
- MixFormerV2: Efficient Fully Transformer Tracking (NeurIPS 2023 · similarity 0.85)
- Target Focused Shallow Transformer Framework for Efficient Visual Tracking (AAAI 2024 · similarity 0.84)
- Track-On: Transformer-based Online Point Tracking with Memory (ICLR 2025 · similarity 0.84)
- Learning Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking (ICML 2024 · similarity 0.83)
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.
