2023 Video Similarity Challenge Codebase

This is the codebase for the 2023 Video Similarity Challenge and the associated dataset.

The Video Similarity Challenge will be featured at the VCDW Workshop at CVPR 2023!

The design and results of the challenge can be found in our paper.

Documentation

The docs folder contains additional documentation for this codebase:

Installation

See installation

Running tests

See testing

Reproducing baseline results

See baseline

Downloading the data

NEW (2024-02-05): the dataset can be found here: dataset

Running evaluations

Descriptor eval

$ ./descriptor_eval.py --query_features vsc_eval_data/queries.npz --ref_features vsc_eval_data/refs.npz --ground_truth vsc_eval_data/gt.csv
Starting Descriptor level eval
...
2022-11-09 17:00:33 INFO     Descriptor track micro-AP (uAP): 0.4754

Matching track eval

$ ./matching_eval.py --predictions vsc_eval_data/matches.csv --ground_truth vsc_eval_data/gt.csv
Matching track segment AP: 0.3650

Baseline results on training set

method	score norm	Descriptor μAP	Matching μAP
SSCD	❌	0.4754	0.3650
SSCD	✔️	0.6499	0.4692
DINO	✔️	0.4402	0.3393
DnS	✔️	0.4129	0.3211

Note: Numbers in italics do not conform to challenge rules

License

The VSC codebase is released under the MIT license.