In this repo, I build a ReID-based tracker that combine both apperance and position information with a Message Passing Neural Network, inspired by the approach presented in Learning a Neural Solver for Multiple Object Tracking.
In Message Passing Neural Network, the node features are ReID embeddings, and the 5-dimensional edge features are as followed:
It is noticeable that in the edge-to-node update, I use the time-aware message passing proposed in Learning a Neural Solver for Multiple Object Tracking.
The dataset used in this experiment is MOT16, which is a benchmark that focus on the pedestrian tracking in the crowded scenes.
The dataset is splitted as followed, as I use leave-one-out cross validation in this experiment.
| Training & Validation Set | Test Set |
|---|---|
| MOT16 - 2 | MOT16 - 1 |
| MOT16 - 4 | MOT16 - 8 |
| MOT16 - 5 | MOT16 - 12 |
| MOT16 - 9 | |
| MOT16 - 10 | |
| MOT16 - 11 | |
| MOT16 - 13 |
The test result are as followed:
| Sequence | MOTA | IDSW |
|---|---|---|
| MOT16-01 | 25.18 | 32 |
| MOT16-08 | 31.29 | 118 |
| MOT16-12 | 45.59 | 31 |
| OVERALL | 33.82 | 181 |
Note: evaluation measures
| Measure | Description |
|---|---|
| MOTA | Multiple object tracker accuracy. |
| IDSW | Identity Switches - Total number of track switches. |