Awesome
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 | :x: | 0.4754 | 0.3650 |
| SSCD | :heavy_check_mark: | 0.6499 | 0.4692 |
| DINO | :heavy_check_mark: | 0.4402 | 0.3393 |
| DnS | :heavy_check_mark: | 0.4129 | 0.3211 |
Note: Numbers in italics do not conform to challenge rules
License
The VSC codebase is released under the MIT license.