Propose training-free, KV-cache reusable video token merging method for long video LLMs that effectively exploits spatio-tempoal locality

Propose a scalable open-vocabulary TAL method that utilizes pseudo-labels from unlabeled videos to expand the semantic context of action localization

Propose an online TAL framework that can detect multiple overlapping action instances in a streaming setting based on a finite-state machine and a novel loss function.

Propose a framework unifying online, semi-online, and offline VIS with a novel training method that bridges the gap between the training and inference stages.

Propose a GNN-based joint detection and tracking method that can track low-scored detections, which were previously filtered out to avoid false positive tracking, by exploiting multi-hop relational features.