论文:Once Detected, Never Lost:
方向:人工标注的优化
Once Detected, Never Lost 流程
- Base Detector
- FSD 修改
- 在离线系统中,可以使用未来帧可能是可以表现更好
- 为了兼顾更多的时序信息,采用了跳帧策略。论文中采样了9帧
[ t − 8 , t − 6 , ⋅ ⋅ ⋅ , t , ⋅ ⋅ ⋅ , t + 6 , t + 8 ] [t − 8, t − 6, · · · , t, · · · , t + 6, t + 8] [t−8,t−6,⋅⋅⋅,t,⋅⋅⋅,t+6,t+8] - 为了防止过拟合,采用了 dropout策略,随机(20%)丢弃一半所采样的帧。
- FSD 修改
- Bidirectional Multi-object Tracking
Track-centric Learning 流程(书接上回:文本核心部分)
- Organizing Track Input
- Multiple-In-Multiple-Out
- Motion-state Agnostic
- Timestamp Encoding
- Sparse Feature Extraction
- Track Features
- Object Features
- Label Assignment
- Track IoU
- Assignment
- Detection Head and Losses
- Classification
- Regression
- Post-processing
- Remove empty predictions.
- Track TTA
- Track Coherence Optimization (TCO)
(先写到这里,后续再补充)