Skip to content

DATASET.md

Latest commit

 

History

History
78 lines (55 loc) · 3.11 KB

DATASET.md

File metadata and controls

78 lines (55 loc) · 3.11 KB

Datasets Documentation

Download

The following datasets are currently supported, they can be downloaded from the respective link:

  • Animals
    • ChimpAct: Chimpanzees in the Leibzig Zoo.
    • AnimalTrack: Diverse selection of 10 common animal categories.
    • PanAf500: Camer trap videos of chimpanzees in their natural environment.
  • Persons
    • DanceTrack: Dancers with unifrom appearance and diverse motion.
    • SportsMot: Athletes in diverse sport scenes.
  • Vehicles
    • UAVDT: Vehicles in complex scenes filmed with drones.
    • BDD100k: Driving videos with multiple object classes.

Processing

After downloading, the datasets need to be unpacked and assembled in the data folder. The goal is to bring them into the standard dancetrack format:

data/
├── dataset/
│   ├── train/
│   │   ├── seq1/
│   │   │   ├── img1/
│   │   │   │   ├── 00001.jpg
│   │   │   │   └── 00002.jpg
│   │   │   └── gt (optional)/
│   │   │       └── gt.txt
│   │   └── seq2/
│   └── test/

The gt.txt contains the annotations in the following format. The box coordinates are in tlwh format and in absolute pixel values. The frame IDs start with 1.

<frame_id>, <track_id>, <box_left>, <box_top>, <box_right>, <box_bottom>, 1, 1, 1

No Processing necessary

These datasets dont need to be processed. They are already in the correct format.

  • DanceTrack
  • SportsMOT

Processing according to Dataset

These datasets need to be processed according to their dataset website.

  • ChimpAct

Custom Processing

These datasets need to be processed with the respective scripts in ./datasets. The script convert the model into dancetrack format and add the auxiliary seqmap and seqinfo files for TrackEval.

  • BFT
  • AnimalTrack
  • GMOT-40-Animal
  • PanAf500

Special Case: BDD100k & UAVDT

These dataset have special characteristics. BDD100k is handled separately and can just be used as is. The results are also saved in a special format. UAVDT has some additional ignore regions, which also need to be supplied in addition to the ground truth.

Adding Custom Dataset

A custom dataset can be added by putting it into the data folder in the dancetrack format and adding its paths and names to the dataset configs in run.py.

dataset_presets = {<dataset name>: {"DATASET": <dataset folder>, "IMG_DIR": <image dir>, "DETECTOR.TEXT_PROMPT": <detection text prompts>}

The dataset can then be used by specifying --dataset <dataset name> when calling run.py.