JDE:Towards Real-Time Multi-object Tracking
One position (x, y) in the output(feature) presents several anchors (w, h ration and different scales). But only one embedding is response for these anchors.
Code:https://github.com/Zhongdao/Towards-Realtime-MOT
Embeddings from multiple scales shares the same space, and association across scales is feasible.
FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking
Code:https://github.com/ifzhang/FairMOT
CenterTrack:Tracking Objects as Points
Code:https://github.com/xingyizhou/CenterTrack
DEFT: Detection Embeddings for Tracking
Combine strengths of the first two algorithms
- Backbone : Center Net
- Loss: Similarity Metrix
- Time sequence: like Siamese Network and buffer of previous frames
- Only one stage : no process for similarity calculation