Tracking Every Thing in the Wild

Current multi-category Multiple Object Tracking (MOT) metrics use class labels to group tracking results for per-class evaluation. Similarly, MOT methods typically only associate objects with the same predictions. These two prevalent strategies in implicitly assume that classification performance is near-perfect. However, this far from case recent large-scale datasets, which contain large numbers of classes many rare or semantically similar categories. Therefore, resulting inaccurate leads sub-optimal and inadequate benchmarking trackers. We address these issues by disentangling tracking. introduce a new metric, Track Every Thing Accuracy (TETA), breaking measurement into three sub-factors: localization, association, classification, allowing comprehensive even under classification. TETA also deals challenging incomplete annotation problem datasets. further tracker (TETer), performs association using Class Exemplar Matching (CEM). Our experiments show evaluates trackers more comprehensively, TETer achieves significant improvements on datasets BDD100K TAO compared state-of-the-art.