Research map / Detection and segmentation / Tracking
Beyond Accuracy: Tracking more like Human via Visual Search
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 CPDTrack, a single-object tracker built on a central-peripheral dichotomy motion model with a temporal information query, plus the STDChallenge benchmark and a human visual Turing test for long-term and global-instance tracking.
Problem
Existing visual object trackers rely on local cropping around the previous target position and fail to match human visual search under spatio-temporal discontinuities (target absence and shot cuts) in long-term and global-instance tracking.
Approach
A motion model inspired by the central-peripheral dichotomy: the previous-frame box defines a Gaussian-acuity central crop with dynamically sized context, while the whole frame is resized as peripheral vision; a temporal information query provides top-down control via cross-attention over past frame tokens.
Key results
- CPDTrack reaches 84.2 N-PRE and 73.3 PRE on STDChallenge, 1.7% and 1.4% above the second-best tracker (SeqTrack/MixViT).
- On VideoCube R-OPE, CPDTrack achieves 82.9 N-PRE and 75.3 robustness, outperforming the second-ranked machine on all metrics.
- CPDTrack achieves the highest error consistency with humans (kappa 0.167) on the STDChallenge-Turing test, versus 0.159 for the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CPDTrack instead of the second-ranked tracker | STDChallenge Benchmark trackers · STDChallenge Benchmark | N-PRE: 81.9 -> 84.2 | Helped |
| use CPDTrack instead of the second-ranked tracker | VideoCube trackers · VideoCube | N-PRE: 76.9 -> 82.9 | Helped |
| add information query (ablation) | CPDTrack · STDChallenge Benchmark | N-PRE: 83.8 -> 84.2 | Helped |
| add information query (ablation) | CPDTrack · VideoCube | N-PRE: 82.7 -> 82.9 | Helped |
| combine central and peripheral vision rather than use central vision only (ablation) | CPDTrack · STDChallenge Benchmark | N-PRE: 81.5 -> 84.2 | Helped |
| combine central and peripheral vision rather than use central vision only (ablation) | CPDTrack · VideoCube | N-PRE: 77.7 -> 82.9 | Helped |
| combine central and peripheral vision rather than use peripheral vision only (ablation) | CPDTrack · STDChallenge Benchmark | N-PRE: 80.7 -> 84.2 | Helped |
| combine central and peripheral vision rather than use peripheral vision only (ablation) | CPDTrack · VideoCube | N-PRE: 80.6 -> 82.9 | Helped |
| use Gaussian acuity for central vision rather than traditional local crop (ablation) | CPDTrack · STDChallenge Benchmark | N-PRE: 83.0 -> 84.2 | Helped |
| use Gaussian acuity for central vision rather than traditional local crop (ablation) | CPDTrack · VideoCube | N-PRE: 84.3 -> 82.9 | Hurt |
Surprising findings
- CPDTrack underperforms on LaSOT, which the authors attribute to limited STDChallenge scenarios and differing ground-truth interpretations.
- Trackers' performance varies strongly with STD while humans remain stable across difficulty levels.
Limitations
- CPDTrack underperforms on LaSOT due to limited STDChallenge scenarios and ground-truth interpretation differences.
- Human visual Turing test uses only five participants on 16 videos, a small-N design.
Experimental setup
Datasets: VideoCube LaSOT GOT-10k TrackingNet VOTLT2019
Benchmarks: STDChallenge VideoCube LaSOT GOT-10k
Metrics: N-PRE PRE SUC AUC P Robust error consistency recovery success rate
Baselines: OSTrack SeqTrack MixViT STARK KeepTrack Ocean SuperDiMP PrDiMP DiMP SiamRCNN GlobalTrack TransT+InT
Builds on: OSTrack MAE ViT SeqTrack MixViT
Closest papers by meaning
- Interactive Tracking: A Human-in-the-Loop Paradigm with Memory-Augmented Adaptation (CVPR 2026 · similarity 0.88)
- A Causal And-Or Graph Model for Visibility Fluent Reasoning in Tracking Interacting Objects (CVPR 2018 · similarity 0.86)
- SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking (CVPR 2025 · similarity 0.85)
- ODTrack: Online Dense Temporal Token Learning for Visual Tracking (AAAI 2024 · similarity 0.85)
- VastTrack: Vast Category Visual Object Tracking (NeurIPS 2024 · similarity 0.84)
- Video Diffusion Models Excel at Tracking Similar-Looking Objects Without Supervision (NeurIPS 2025 · similarity 0.84)
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.