TL;DR We introduce Track-On-R, a point tracker adapted to real-world videos using verifier-guided pseudo-labeling. The verifier identifies reliable predictions from multiple trackers and enables training on unlabeled real videos, significantly improving tracking robustness across diverse real-world benchmarks while using fewer real-world videos.

Domain Gap in Point Tracking

Different trackers fail under different conditions and frames, highlighting the limitations of single-teacher pseudo-labeling.

Synthetic Training

Most point trackers are trained on synthetic videos with labels, which do not fully capture real-world appearance and motion complexity.

Single-Teacher Limitation

Pseudo-labeling with one teacher propagates its errors directly, making real-world adaptation brittle.

Verifier Motivation

Since different trackers succeed under different conditions, we learn a verifier to identify the most reliable prediction.

Verifier-Guided Real-World Adaptation

Verifier-guided pseudo-labeling pipeline

Left: Candidate trajectories produced by multiple off-the-shelf trackers. Right: A verifier selects the most reliable prediction to generate pseudo-labels for adapting a tracker to real-world videos.

Candidate trajectories. Multiple off-the-shelf trackers produce alternative predictions for the same query point.
Verifier scoring. A lightweight verifier evaluates the reliability of each candidate trajectory using local visual evidence and temporal consistency.
Verifier-guided pseudo-labeling. The most reliable trajectory is selected and used as pseudo-labels for real-world adaptation.
Track-On-R. Using these pseudo-labels, we fine-tune the base tracker on unlabeled real-world videos, resulting in Track-On-R, a model adapted to real-world motion and appearance variations.

Results

Benchmark Comparison

Model	EgoPoints		RoboTAP		Kinetics		DAVIS
Model	δ_avg	OA	δ_avg	OA	δ_avg	OA	δ_avg	OA
BootsTAPNext	33.6	89.5	75.0	88.7	70.6	87.4	78.5	91.2
CoTracker3	54.0	84.4	78.8	90.8	68.5	88.3	76.3	90.2
AllTracker	62.0	87.1	80.9	92.2	69.3	89.1	77.0	88.7
Track-On-R (Ours)	67.3	90.2	82.6	94.0	71.0	90.5	80.3	92.5

Qualitative Results

DAVIS

Kinetics

RoboTAP

Credits / Acknowledgments

Datasets. Qualitative examples shown on this page are drawn from publicly available tracking datasets, including LaSOT and AnimalTrack. We thank the authors of these datasets for collecting and releasing them.

Compared Trackers. We also acknowledge the authors of the tracking methods used in our comparisons, including BootsTAPNext, AllTracker, and CoTracker3, for making their models publicly available.

Previous Work

This work builds on our previous research on online point tracking.

Track-On2: Enhancing Online Point Tracking with Memory
Görkay Aydemir, Weidi Xie, Fatma Güney
under submission, 2025

BibTeX | arXiv | Project Page

@article{aydemir2025trackon2,
  title   = {Track-On2: Enhancing Online Point Tracking with Memory},
  author  = {Aydemir, G\"orkay and Xie, Weidi and G\"uney, Fatma},
  journal = {arXiv preprint arXiv:2509.19115},
  year    = {2025}}

Track-On: Transformer-based Online Point Tracking with Memory
Görkay Aydemir, Xiongyi Cai, Weidi Xie, Fatma Güney
ICLR, 2025

BibTeX | arXiv | Project Page

@inproceedings{aydemir2025trackon,
  title     = {Track-On: Transformer-based Online Point Tracking with Memory},
  author    = {Aydemir, G\"orkay and Cai, Xiongyi and Xie, Weidi and G\"uney, Fatma},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2025}
}