January 2020
tl;dr: Summary of the main idea.
Overall impression
There is a series of paper on recurrent single stage method for object detection. The main idea is to add RNN layer directly on top of the entire image feature.
The conversion of bbox to heatmap is also another example of transforming unstructured information to pseudo-image.
K = 6 frames
Key ideas
- Using historical visual semantics to improve tracking. (Although there is only one object in the image)
- When assigning detection to tracklet, use IoU distance between the current detection and the mean of its short-term history of validated detections.
- Training is multi-staged. First train the network on single image detection.
- Three inputs to LSTM:
- 4096-d feature vector for the entire image
- heatmap from detection of the current frame
- Output from last time-step
Technical details
- Evaluation metrics of tracking: Success Plots, accuracy (success ratio) vs IoU thresholds
- OPE (one pass evaluation): all frames
- TRE (temporal robustness evaluation): random frame as starting frame
- SRE (spatial robustness evaluation): jittered GT bbox
- Under the same frames of GT, more video with sparse annotation is more useful than fewer video with dense annotation.
- YOLO with Kalman filter (SORT) performs poorly due to fast motions, occlusions, and therefore occasionally poor detections.
Notes