Related Experiment Video
Updated: Jul 11, 2026
08:06
Eye Movement Monitoring of Memory
Published on: August 15, 2010
15.3K
Track-On2: Enhancing Online Point Tracking With Memory
Summary
Track-On2 is a new transformer model for online long-term point tracking. It achieves state-of-the-art results by using causal processing and a memory mechanism for robust tracking in real-time applications.
Area of Science:
- Computer Vision
- Machine Learning
Background:
- Long-term point tracking is challenging due to appearance changes, motion, and occlusion.
- Online tracking requires frame-by-frame processing for real-time applications.
Purpose of the Study:
- To develop an efficient and effective online long-term point tracking model.
- To improve upon existing tracking methods by addressing limitations in handling appearance changes and occlusions.
Main Methods:
- Introduced Track-On2, a transformer-based model extending the prior Track-On model.
- Implemented causal frame processing with a memory mechanism to maintain temporal coherence.
- Utilized coarse patch-level classification followed by refinement during inference.
- Investigated synthetic training strategies to enhance temporal robustness.
Main Results:
- Track-On2 achieved state-of-the-art performance on five synthetic and real-world benchmarks.
- Outperformed prior online trackers and strong offline methods.
- Demonstrated effectiveness in handling significant appearance changes, motion, and occlusion.
Conclusions:
- Causal, memory-based architectures trained on synthetic data are scalable solutions for real-world point tracking.
- Track-On2 offers improved performance and efficiency for online long-term tracking applications.
