cli: add download command for benchmark datasets

[SkalskiP]

Description

Add a trackers download CLI subcommand that lets users download benchmark datasets from a GCP bucket. Supports downloading full datasets or individual splits, with or without video frames. Datasets: MOT17, SportsMOT, SoccerNet Tracking, DenseTrack.

Proposed CLI interface

# List what's available
trackers download --list
# Download full dataset
trackers download mot17
# Download specific split, annotations only
trackers download mot17 --split train --annotations-only
# Custom output path
trackers download sportsmot --split val -o ./datasets

Tasks

• Research and document the structure of each dataset (MOT17, SportsMOT, SoccerNet Tracking, DenseTrack): available splits, contents per split (frames, GT annotations, detections), directory layout.
• Prepare zip files for each dataset/split/variant combination and upload them to a GCP bucket.
• Create a dataset manifest (mapping dataset + split + variant to GCP URLs and metadata) inside the codebase.
• Register the download subcommand in trackers/scripts/__main__.py.
• Implement the --list flag to display available datasets, splits, and content types in a formatted table.
• Implement download logic: resolve URL from manifest, download zip with progress bar, extract to output directory.
• Handle edge cases: existing files (skip/overwrite), network errors, interrupted downloads.
• Test the command end-to-end for each dataset and variant.

[omkar-334]

hey @SkalskiP , I'd like to work on this. i'll get started by looking into the datasets and documenting the structure & splits.
I think the above mentioned PR is by an AI bot

[SkalskiP]

Yeah, I closed the PR above because it was auto-generated and did not meet the basic requirements for this task.

The first step is to plan which .zip files we will need for each dataset task. I’d prefer you work on one task at a time. You are already quite far along with #244. How about we finish #244 first, and if #246 is still open afterward, you can pick that up as well?

[SkalskiP]

Dataset	Frames Train	Frames Test	Frames Valid	GT Train	GT Test	GT Valid	Detections Train	Detections Test	Detections Valid
MOT17	Yes	Yes	No★	Yes	No	No★	Yes	Yes	No★
SportsMOT	Yes	Yes	Yes	Yes	No	Yes	No	No	No
SoccerNet Tracking	Yes	Yes	No	Yes	Yes	No	Yes	Yes	No
DenseTrack	Yes	Yes	Yes	Yes	No	Yes	No	No	No

★ MOT17 has no official validation split. Validation is typically created from the train split.

@tstanczyk95 is this correct summary of all 4 datasets?

[tstanczyk95]

Dataset Frames Train Frames Test Frames Valid GT Train GT Test GT Valid Detections Train Detections Test Detections Valid
MOT17 Yes Yes No★ Yes No No★ Yes Yes No★
SportsMOT Yes Yes Yes Yes No Yes No No No
SoccerNet Tracking Yes Yes No Yes Yes No Yes Yes No
DenseTrack Yes Yes Yes Yes No Yes No No No
★ MOT17 has no official validation split. Validation is typically created from the train split.

@tstanczyk95 is this correct summary of all 4 datasets?

@SkalskiP A few remarks here:

• MOT17 provides public detections for the train set, but in practice these are not used today. Everyone goes for private detector (provided by themselves or by a source different than MOT17)
• SoccerNet tracking provides oracle detections and ground truths for train and test set.
• SoccerNet tracking, does not have a validation set. Only train set, test set, challenge 2022 set and challenge 2023 sets. However, evaluation on the challenges 2022 and 2023 is not commonly used for algorithm evaluation and it is currently disabled.

[omkar-334]

looks like i can take this up now @SkalskiP ?

[SkalskiP]

Yes! You feel you know what to do?

[JVSCHANDRADITHYA]

Hi @SkalskiP, I’ve reviewed the datasets and I think I can take this up.

[omkar-334]

Yes! You feel you know what to do?

Yes, thanks for the help!

[JVSCHANDRADITHYA]

@SkalskiP @tstanczyk95
I was able to find and download MOT17, SportsMOT, and SoccerNet Tracking, but I’m unable to find any official release or download link for DenseTrack, so I feel like I’m missing something. I did come across the paper “DenseTrack: Drone-based Crowd Tracking via Density-aware Motion-appearance Synergy” (arXiv:2407.17272), but it uses the DroneCrowd dataset and benchmarks on multiple other datasets, which seems to be something else. Could you please clarify which dataset is referred to as DenseTrack here and point me to the correct source if available? Thanks.

[JVSCHANDRADITHYA]

Datasets reviewed :

• MOT17 : official MOTChallenge release
• SportsMOT: official OneDrive release
• SoccerNet Tracking : official sn-tracking repository / pip package download available
• DenseTrack : source unclear, request for clarification. [ clarification : DanceTrack ]
• DanceTrack : official release on hugging face

[SkalskiP]

Hi @omkar-334 and @JVSCHANDRADITHYA, that was a typo. I meant DanceTrack, not DenseTrack. I’m sorry for the confusion.

@JVSCHANDRADITHYA could you please describe the steps required to download the dataset and share the links to the original releases?

[JVSCHANDRADITHYA]

@SkalskiP Thanks for the clarification — got it, DanceTrack.

• MOT17: Official MOTChallenge release
  https://motchallenge.net/data/MOT17/
  Download via OneDrive → extract train/test sequences (frames, GT for train, detections for both).

• SportsMOT: Official OneDrive release
  https://onedrive.live.com/?redeem=aHR0cHM6Ly8xZHJ2Lm1zL3UvcyFBdGplTHE3WW5ZR1JnUVBtdUFrd1dobmRPYzQxP2U9SXRKYWFh&cid=91819DD8AE2EDED8&id=91819DD8AE2EDED8%21131&parId=91819DD8AE2EDED8%21130&o=OneUp

• SoccerNet Tracking: Official pip / GitHub release
  https://github.com/SoccerNet/sn-tracking
  Installed via pip install sn-tracking, which downloads and prepares the dataset programmatically.

I’ll have MOT17 and SoccerNet Tracking fully documented and initial implentations (structure, splits, ZIP plan) by EOD. SoccerNet is straightforward since it’s a pip-based workflow.

One Quick clarification: for SoccerNet Tracking, is it acceptable to invoke the official pip-based download process from the CLI instead of mirroring it in a GCP bucket?