Research map / Detection and segmentation / Tracking
Explicit Visual Prompts for Visual Object 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 EVPTrack, a transformer tracker that propagates spatio-temporal tokens across frames and turns them, plus multi-scale template features, into explicit visual prompts, avoiding template-update strategies.
Problem
Visual object tracking must capture target appearance changes using spatio-temporal context, but existing template-updating trackers face the when-and-how-to-update dilemma and lack inter-frame context.
Approach
Uses spatio-temporal tokens carried between consecutive frames, plus multi-scale template features, as explicit visual prompts concatenated with image tokens in a single transformer encoder, instead of updating templates.
Key results
- EVPTrack-384 reaches 72.7% AUC on LaSOT, above the compared trackers, including ARTrack-384 at 72.6%.
- EVPTrack-224 runs at 71 FPS with 21 GFLOPs; EVPTrack-384 has 65 GFLOPs at 28 FPS versus SeqTrack-B384 at 148 GFLOPs and 11 FPS.
- Ablation on LaSOT: explicit visual prompts raise AUC from 69.4 (baseline) to 70.4 for EVPTrack-224.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add spatio-temporal and multi-scale explicit visual prompts (ablation) | Baseline tracker · LaSOT | AUC: 69.4 -> 70.4 | Helped |
| Replace explicit visual prompts with learnable tokens (ablation) | Baseline tracker · LaSOT | AUC: 69.4 -> 69.6 | Helped |
| Use multi-scale prompts (ablation) | Baseline tracker · LaSOT | AUC: 69.4 -> 70.0 | Helped |
| Use only spatio-temporal prompts (ablation) | Baseline tracker · LaSOT | AUC: 69.4 -> 69.7 | Helped |
| Use template token (ablation) | Baseline tracker · LaSOT | AUC: 69.4 -> 69.2 | Hurt |
| Use single-scale 14x14 prompts instead of multi-scale prompts (ablation) | EVPTrack prompt design · LaSOT | AUC: 70.0 -> 69.6 | Hurt |
| Use single-scale 16x16 prompts instead of multi-scale prompts (ablation) | EVPTrack prompt design · LaSOT | AUC: 70.0 -> 68.9 | Hurt |
| Use single-scale 18x18 prompts instead of multi-scale prompts (ablation) | EVPTrack prompt design · LaSOT | AUC: 70.0 -> 69.7 | Hurt |
| Change training sampling from M=4, N=8 to M=8, N=4 (ablation) | EVPTrack-224 · LaSOT | AUC: 70.4 -> 69.9 | Hurt |
| Change training sampling from M=4, N=8 to M=2, N=16 (ablation) | EVPTrack-224 · LaSOT | AUC: 70.4 -> 69.2 | Hurt |
Surprising findings
- Replacing the explicit prompts with learnable implicit tokens gives a smaller gain (69.6 AUC) than multi-scale or spatio-temporal prompts alone, and using template tokens as propagated tokens lowers AUC (69.2).
- Increasing training sequence length from 8 to 16 frames decreased LaSOT AUC by 1.2%, attributed to only two videos per iteration.
Limitations
- Limited GPU memory restricts the training video sequence length to 4-8 frames, which limits exploitation of long-range temporal information.
- Lower input resolution (224) hurts performance on LaSOT_ext, which the authors attribute to insufficient background context.
Experimental setup
Datasets: LaSOT GOT-10k TrackingNet COCO
Benchmarks: LaSOT LaSOT_ext GOT-10k UAV123 TrackingNet TNL2K
Metrics: AUC normalized precision precision AO SR0.5 SR0.75 FPS FLOPs params
Baselines: OSTrack-384 SeqTrack-B384 ARTrack-384 STARK SimTrack-B/16 MixFormer-22k TransT AiATrack
Builds on: HiViT MAE OSTrack SeqTrack STARK
Closest papers by meaning
- HIPTrack: Visual Tracking with Historical Prompts (CVPR 2024 · similarity 0.87)
- MIMTrack: In-Context Tracking via Masked Image Modeling (AAAI 2025 · similarity 0.87)
- Efficient Motion Prompt Learning for Robust Visual Tracking (ICML 2025 · similarity 0.87)
- ODTrack: Online Dense Temporal Token Learning for Visual Tracking (AAAI 2024 · similarity 0.86)
- Autoregressive Sequential Pretraining for Visual Tracking (CVPR 2025 · similarity 0.86)
- UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking (CVPR 2026 · similarity 0.86)
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.