Research map / Vision-language and multimodal / Multimodal LLMs

UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 0 citations (OpenAlex)

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 UTPTrack, a one-stream tracker token-pruning framework that jointly prunes search, dynamic template, and static template tokens (with a token-type-aware foreground bonus and text-guided pruning for unified tracking), achieving ~65% token reduction with near-baseline accuracy on OSTrack and SUTrack.

Problem

One-stream Transformer trackers are computationally expensive for real-time deployment, and existing token pruning methods prune the search region or dynamic template in isolation rather than jointly with the static template.

Approach

Attention-guided CTEM modules prune search, dynamic-template, and static-template tokens jointly in a one-stream encoder, scoring each token by attention to the static template's center token; a token-type-aware soft foreground bonus from the target box protects foreground static-template tokens, and a text-guided score extends pruning to language-guided unified tracking.

Key results

• UTPTrack-O256 retains 99.5% of baseline performance while pruning 24.5% of tokens on RGB tracking, and 100.2% at 18.8% pruning.
• UTPTrack-S224 achieves 99.7% of baseline performance with 65.4%-ish vision token reduction at 224 resolution and 100.0% with text guidance on unified ablation; at 384 resolution it reaches 100.5% with 67.5% tokens pruned.
• Pruning reduces MACs by 31.3% (OSTrack-384) and 28.4% (SUTrack-384) while keeping roughly baseline accuracy.

Surprising findings

• Moderate pruning slightly improves accuracy over the unpruned baseline, which the authors attribute to removing noisy redundant tokens acting as a regularizer.
• Injecting language cues into the search region or static template degrades performance, while injecting into the dynamic template alone matches the baseline.

Limitations

• Gains on GPU FPS are modest at some resolutions (e.g., +1 FPS at 256 for OSTrack).
• Text-guided pruning degrades performance on TNL2K, attributed to imprecise language descriptions in long videos.
• Evaluated only on OSTrack and SUTrack base trackers.

Experimental setup

Datasets: TrackingNet LaSOT GOT-10k COCO TNL2K VASTTrack DepthTrack LasHeR VisEvent

Benchmarks: LaSOT LaSOT_ext TrackingNet GOT-10k VOT-RGBD22 LasHeR RGBT234 VisEvent TNL2K OTB99

Metrics: AUC P P_Norm AO SR EAO MSR MPR MACs FPS

Baselines: OSTrack-CE OSTrack-ToMe OSTrack-EViT SUTrack-CE SUTrack-ToMe SUTrack-EViT SUTrack-DynamicViT

Builds on: OSTrack SUTrack CE (candidate elimination) STARK

Closest papers by meaning

• ODTrack: Online Dense Temporal Token Learning for Visual Tracking (AAAI 2024 · similarity 0.88)
• SUTrack: Towards Simple and Unified Single Object Tracking (AAAI 2025 · similarity 0.87)
• Exploring Token Pruning in Vision State Space Models (NeurIPS 2024 · similarity 0.87)
• Less Is More: Token Context-Aware Learning for Object Tracking (AAAI 2025 · similarity 0.87)
• Single-Model and Any-Modality for Video Object Tracking (CVPR 2024 · similarity 0.87)
• MIMTrack: In-Context Tracking via Masked Image Modeling (AAAI 2025 · similarity 0.86)