DatasetHarvester

Easily download and set up Video Object Segmentation (VOS) datasets.

Overview

Dataset	Status	Version(s)	Attributes
SA-V	✅ Semi-Auto	-	-
LVOS	✅ Auto	v1, v2	Long-Term
MOSE	✅ Auto	v1, v2	Complex Scenarios
OVIS	✅ Auto	-	-
BURST	✅ Auto	-	-
PUMaVOS	✅ Auto	-	Partial Masks
VOST	✅ Auto	-	-
Static	✅ Auto	-	Only for Training
DiDi	⚠️ Manual or vot-toolkit	-	Distractor Heavy
VOTS	⚠️ vot-toolkit	-	Annual Challenge
DAVIS	⚠️ external script	2016-2019	-
YouTube-VOS	⚠️ external script	2018-2022	-

Tip

Descriptions, Official Repos & citations in documentation/DATASETS.md. URLs for supported_datasets.yaml and not_supported_datasets.yaml.

Quick installation

Run:

git clone https://github.com/Vujas-Eteph/DatasetHarvester
cd DatasetHarvester
uv sync

Download LVOS, MOSE, OVIS, BURST, PUMaVOS, VOST & Static

Run:

uv run WizHarvester.py

Download Meta's SA-V Dataset

1. Fill this Meta document out to get access to the URLs.
2. Adapt download_SA_V.yaml with the URLs.
3. Adapt the paths in line 44 of download_SA_V.py
4. Run

   cd scripts/SA_V/
   uv run download_SA_V.py

5. More stuff HERE

Download DAVIS & YouTube-VOS

Use this script to download YouTubeVOS and DAVIS.

Manual Downloads

Download and extract DiDi manually.

For the VOTS dataset/benchmark

The vot-toolkit handles the installation automatically, only need to find the correct stack-name

pip install vot-toolkit
vot initialize <stack_name>  --workspace <stack-name>

More info:
Overview

---

Note

Contributions are welcome to help keep this repo up to date. 🤗

Tip

Modifed versions to act as pip packages for local evaluations:
vos-benchmark: https://github.com/hkchengrex/vos-benchmark.git
lvos-api: https://github.com/Vujas-Eteph/lvos-evaluation
mose-api: https://github.com/Vujas-Eteph/MOSE-api