CoTracker3: Simpler and Better Point Tracking by Pseudo-Labeling Real Videos

Figure 1. Scaling point trackers with unsupervised videos. Left: We compare our CoTracker3 to other point trackers. Each model is pre-trained on synthetic data (Kubric) and then fine-tuned on real videos using our simple unsupervised training protocol. Our model and training protocol outperform SoTA by a large margin using only 0.1% of the real training data. Right: CoTracker3 is robust to occlusions.