Grounding-Tracking-Integration

In this paper, we study tracking by language that localizes the target box sequence in a video based on query. We propose framework called GTI decomposes problem into three sub-tasks: Grounding, Tracking, and Integration. The sub-task modules operate simultaneously predict frame-by-frame. “Grounding” predicts referred region directly from “Tracking” history of grounded regions previous frames. “Integration” generates final predictions synergistically combining grounding tracking. With “integration” task as key, explore how to indicate quality each frame achieve desired mutually beneficial combination. To end, an “RT-integration” method defines two scores guide integration: 1) R-score represents Region correctness whether prediction accurately covers target, 2) T-score Template provides informative visual cues improve future present our real-time implementation with proposed RT-integration, benchmark LaSOT Lingual OTB99 highly promising results. Moreover, produce disambiguated version queries facilitate studies.