DATASET.md

Data Preparation

🔗 For all the datasets we used in our experiments, you can access them from the following public link:

• DanceTrack
• SportsMOT
• MOT17
• CrowdHuman

Generate GT files

For the MOT17 and CrowdHuman datasets, you’ll need to use the provided script to convert their ground truth files to the format we require:

• For MOT17: gen_mot17_gts.py
• For CrowdHuman: gen_crowdhuman_gts.py

📌 You need to modify the paths in the script according to your requirements.

File Tree

<DATADIR>/
  ├── DanceTrack/
  │ ├── train/
  │ ├── val/
  │ ├── test/
  │ ├── train_seqmap.txt
  │ ├── val_seqmap.txt
  │ └── test_seqmap.txt
  ├── SportsMOT/
  │ ├── train/
  │ ├── val/
  │ ├── test/
  │ ├── train_seqmap.txt
  │ ├── val_seqmap.txt
  │ └── test_seqmap.txt
  ├── MOT17/
  │ ├── images/
  │ │ ├── train/     # unzip from MOT17, you need to delete some files, see the below the instructions
  │ │ └── test/      # unzip from MOT17
  │ └── gts/
  │   └── train/     # generate by ./data/gen_mot17_gts.py
  └── CrowdHuman/
    ├── images/
    │ ├── train/     # unzip from CrowdHuman
    │ └── val/       # unzip from CrowdHuman
    └── gts/
      ├── train/     # generate by ./data/gen_crowdhuman_gts.py
      └── val/       # generate by ./data/gen_crowdhuman_gts.py

⚠️ Since each video sequence in MOT17 is stored three times, for each training video sequence, you should delete the other two sequences to achieve deduplication. For instance, in my experiment, I only retained the ‘xxx-DPM’ sequences.”

Q & A

• Q: Lack the val_seqmap.txt file of SportsMOT?
  A: Refer to The 'val_seqmap.txt' file of SportsMOT dataset · Issue #13