Research map / Architectures and efficiency / Attention and transformers
Target Focused Shallow Transformer Framework for Efficient Visual Tracking
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 a target-focused shallow transformer tracker that uses a Gaussian-inspired 2D binary mask in self-attention to suppress background features, with a siamese backbone and center-ness prediction head, reporting OTB100/UAV123 results at about 34 FPS.
Problem
Standard self-attention in transformer trackers mixes foreground and background information and is computationally expensive, causing distraction by background clutter and slow tracking speed.
Approach
Modify self-attention with a Gaussian map-inspired 2D binary mask that focuses attention on the target region in feature maps, combined with a shallow transformer encoder-decoder inside a siamese tracker and a classification/regression/center-ness head.
Key results
- On OTB100 the tracker reports 69.6% success and 90.3% precision, versus TransT's 69.4% and 89.9%.
- On UAV123 it reports 70.2% success and 87.2% precision, versus TransT's 68.1% and 87.6% precision and SeqTrack's 68.5% and 89.1%.
- The tracker runs at approximately 34 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the target-focused shallow Transformer framework instead of TransT | visual tracking · OTB100 | success score (%): 69.4 -> 69.6 | Helped |
| use the target-focused shallow Transformer framework instead of TransT | visual tracking · UAV123 | success score (%): 68.1 -> 70.2 | Helped |
| use the target-focused shallow Transformer framework instead of TransT | visual tracking · OTB100 | precision score (%): 89.9 -> 90.3 | Helped |
| use the target-focused shallow Transformer framework instead of TransT | visual tracking · UAV123 | precision score (%): 87.6 -> 87.2 | Hurt |
| use the target-focused shallow Transformer framework instead of SeqTrack | visual tracking · OTB100 | success score (%): 68.3 -> 69.6 | Helped |
| use the target-focused shallow Transformer framework instead of SeqTrack | visual tracking · UAV123 | success score (%): 68.5 -> 70.2 | Helped |
| use the target-focused shallow Transformer framework instead of SeqTrack | visual tracking · OTB100 | precision score (%): 89.1 -> 90.3 | Helped |
| use the target-focused shallow Transformer framework instead of SeqTrack | visual tracking · UAV123 | precision score (%): 89.1 -> 87.2 | Hurt |
Surprising findings
- On UAV123 the proposed tracker has lower precision (87.2%) than both TransT (87.6%) and SeqTrack (89.1%).
Limitations
- Evaluated only on OTB100 and UAV123 in this version; GOT10k, LaSOT, and TrackingNet are deferred to future work.
- Method details such as the mask construction and training setup are sparsely described.
- Does not clearly beat baselines on UAV123 precision.
Experimental setup
Benchmarks: OTB100 UAV123
Metrics: success score precision score FPS
Baselines: TransT SeqTrack
Builds on: SiamFC TransT
Closest papers by meaning
- Compact Transformer Tracker with Correlative Masked Modeling (AAAI 2023 · similarity 0.89)
- Transformer Tracking With Cyclic Shifting Window Attention (CVPR 2022 · similarity 0.87)
- Learning Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking (ICML 2024 · similarity 0.84)
- Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking (CVPR 2025 · similarity 0.84)
- Is Multiple Object Tracking a Matter of Specialization? (NeurIPS 2024 · similarity 0.84)
- PADRe: A Unifying Polynomial Attention Drop-in Replacement for Efficient Vision Transformer (ICLR 2025 · 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.