GOT-JEPA: Generic Object Tracking with Model Adaptation and Occlusion Handling using Joint-Embedding Predictive Architecture

The article introduces GOT-JEPA, a model-predictive pretraining framework designed to improve generic object tracking by enhancing generalization and occlusion handling. It employs a joint-embedding predictive architecture where a teacher predictor generates pseudo-tracking models from clean frames, while a student predictor learns from corrupted frames, providing stable supervision. Additionally, the framework includes OccuSolver, which refines visibility states for better occlusion perception, resulting in improved performance across seven benchmarks, making it significant for practitioners focusing on robust object tracking in dynamic environments.