One More Check: Making “Fake Background” Be Tracked Again

• DOI: 10.1609/aaai.v36i2.20045
• Publication Date: 2022-07-04T10:34:19Z
• Authors: Chao Liang; Zhipeng Zhang; Xue Zhou; Bing Li; Weiming Hu
• Keywords: Bounding overwatch; Minimum bounding box; BitTorrent tracker; Code (set theory); Motion blur

The one-shot multi-object tracking, which integrates object detection and ID embedding extraction into a unified network, has achieved groundbreaking results in recent years. However, current trackers solely rely on single-frame detections to predict candidate bounding boxes, may be unreliable when facing disastrous visual degradation, e.g., motion blur, occlusions. Once target box is mistakenly classified as background by the detector, temporal consistency of its corresponding tracklet will no longer maintained. In this paper, we set out restore boxes misclassified ``fake background'' proposing re-check network. network innovatively expands role from data association forecasting effectively propagating previous tracklets frame with small overhead. Note that propagation are yielded an independent efficient search, preventing model over-relying results. Eventually, it helps reload repair broken tracklets. Building strong baseline CSTrack, construct new tracker achieve favorable gains 70.7 ➡ 76.4, 70.6 76.3 MOTA MOT16 MOT17, respectively. It also reaches state-of-the-art IDF1 performance. Code released at https://github.com/JudasDie/SOTS.