Video Object Verification via Meta-learning

Abstract

Performance of a long term object tracker relies on the object detection accuracy. Although several object detectors are proposed in the literature, robustness to target disappearances and reappearances is still a challenging problem. To deal with this problem, we propose an inference pipeline that integrates an object detector with a meta-learner, both locally trained. This is achieved by replacing the head classification layer of the object detector by a meta-learner that also enables verification of the target. In particular, Mask R-CNN object detector is integrated with SDNet trained end-to-end for object tracking. Improvement achieved by MAML++ meta learner trained as a classifier is also evaluated. Numerical results reported on VOT2020-LT long term video dataset demonstrate that both SDNet and MAML++ meta-learners improve the detection accuracy for unseen object classes. Moreover verification by SDNET provides 7% increase on detection of target disappearance and reappearance frames.