Lekim89/mot17-parquet

MOT17 – Normalized Parquet Dataset

Multi-Object Tracking Benchmark (MOTChallenge) in a de-duplicated, Parquet-native layout.

Structure

mot17-parquet/
├── data/
│   ├── frames/          # frame metadata (one row per image)
│   │   ├── train-00000-of-00001.parquet
│   │   ├── test-00000-of-00001.parquet
│   │   └── ablation-00000-of-00001.parquet
│   ├── gt/              # ground-truth annotations (train only)
│   │   ├── train-00000-of-00001.parquet
│   │   └── ablation-00000-of-00001.parquet
│   ├── detections/      # public detections, one subfolder per detector
│   │   ├── dpm/
│   │   ├── frcnn/
│   │   └── sdp/
│   └── seqinfo/
│       └── seqinfo.parquet
├── images/
│   ├── train/
│   │   ├── MOT17-02/img1/000001.jpg
│   │   └── ...
│   └── test/
│       ├── MOT17-01/img1/000001.jpg
│       └── ...
└── tools/
    └── export_motchallenge_layout.py

Splits

Split	Description
train	Full training sequences (MOT17-02, 04, 05, 09, 10, 11, 13)
test	Official test sequences (MOT17-01, 03, 06, 07, 08, 12, 14) – no GT
ablation	Second half of each training sequence (ByteTrack convention)

Key design decisions

• Images stored once per sequence (not duplicated across DPM/FRCNN/SDP).
• Detections separated by detector type: dpm, frcnn, sdp.
• Ablation split = second half of each train sequence (frame >= seqLength // 2 + 1).
• Original MOTChallenge column semantics preserved with descriptive names.

Table schemas

frames

Column	Type	Description
sequence	str	e.g. MOT17-02
frame	int	1-based frame number
image_path	str	relative path to image file
width	int	frame width
height	int	frame height
fps	int	sequence frame rate
seq_length	int	total frames in sequence

gt

Column	Type	Description
sequence	str	e.g. MOT17-02
frame	int	1-based frame number
track_id	int	identity
bbox_left	float	bounding box left
bbox_top	float	bounding box top
bbox_width	float	bounding box width
bbox_height	float	bounding box height
conf	float	confidence (1 for GT)
class_id	int	object class
visibility	float	occlusion ratio

detections

Column	Type	Description
sequence	str	e.g. MOT17-02
detector	str	DPM, FRCNN, or SDP
frame	int	1-based frame number
bbox_left	float	bounding box left
bbox_top	float	bounding box top
bbox_width	float	bounding box width
bbox_height	float	bounding box height
score	float	detection confidence

Usage

from datasets import load_dataset

# Load GT for train
gt = load_dataset("Lekim89/mot17-parquet", "gt", split="train")

# Load FRCNN detections for ablation
frcnn = load_dataset("Lekim89/mot17-parquet", "detections-frcnn", split="ablation")

# Load frame metadata
frames = load_dataset("Lekim89/mot17-parquet", "frames", split="train")

Selective image download

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="Lekim89/mot17-parquet",
    repo_type="dataset",
    allow_patterns=[
        "images/train/MOT17-02/**",
        "data/gt/train-*.parquet",
        "data/detections/frcnn/train-*.parquet",
    ],
)

Citation

@article{MOT16,
  title={MOT16: A Benchmark for Multi-Object Tracking},
  author={Milan, A. and Leal-Taix{\'e}, L. and Reid, I. and Roth, S. and Schindler, K.},
  journal={arXiv:1603.00831},
  year={2016}
}