- Release our custom dataset.
- Finalize scripts for each experiment setup.
- Update additional experimental results.
- 06/08/2024: Our custom detector/motion model weights are released!
- 06/04/2024: Initial version of code released.
In Object Tracking, "ID Switch" refers to a problem where the identities of two or more objects are swapped during tracking as they overlap. This is a common issue in Multi-Object Tracking, and some research addresses this problem using appearance-based matching.
Then, how well would the MOT models perform in environments where objects are identical in appearance with fast motion?
From this motivation, this project aims to 1) measure the performance of several current MOT works, in conditions where object appearance is completely identical) and 2) identify ways to improve performance. For this experiment, we have chosen Yabawi (the Shell Game), specifically the Matryoshka mini-game from Nintendo Mario Party, as a suitable task. We conduct several experiments based on ByteTrack to seek improvments in performance.
🪆 Sohyun Yoo (YAI 12th) - Baseline (DeepSORT, SparseTrack) | Main (Re-ID, Depth Tracker) Experiments
🪆 Jian Kim (YAI 12th) - Custom Dataset Construction
🪆 Junghyun Park (YAI 12th) - Related Works / Main Experiments (Mixed Tracker)
🪆 Gun Jegal (YAI 12th) - Baseline Experiments (FairMOT)
🪆 Kyunghoon Jung (YAI 12th) - Yolov8 Training / Baseline (ByteTrack) | Main Experiments (LSTM Tracker)
🪆 Jimin Lee (YAI 11th) - Baseline Experiments (MOTRv2)
We release our annotated Task Dataset:
- TBA
To set up the project, follow these steps:
git clone https://github.com/cygbbhx/kslr.git
We follow the setup from ByteTrack.
You can download our pretrained detector and LSTM motion model from here. Make sure you put the weights in pretrained/ folder.
To run tracking, use the following command:
bash ./scripts/run_ours.sh--wandb: run sweep based on the provided configuration. (Note: You should modify the configuration according to experiment settings.)
| model | IDF1 | IDs | MOTA |
|---|---|---|---|
| DeepSORT | 27.9 |
62 |
56.0 |
| MOTRv2 | 21.7 |
43 |
45.1 |
| ByteTrack | 44.4 |
60 |
91.3 |
| SparseTrack | 40.5 |
60 |
91.6 |
| method | IDF1 | IDs | MOTA |
|---|---|---|---|
| ByteTrack | 44.4 |
60 |
91.3 |
| ByteTrack + Cascading | 53.12 |
58 |
91.7 |
| ByteTrack + Cascading + Re-ID | 50.8 |
49 |
86.5 |
| ByteTrack + Cascading + Re-ID + reset KF | 55.11 |
46 |
96.12 |
- Note. We experiment various thresholds for each method and choose the best results.
Our code is built heavily upon ByteTrack. Our sincere thanks for their wonderful works and contributions.
Note that this is a toy project done in a student organization. If you find any solutions to the problem, you are more than welcome to contact us. If you have any questions, please contact cygbbhx@yonsei.ac.kr.