Repository navigation
Secondsight v0.2.0 is the first tagged release. It turns the original research notebook into a tested, installable package and ships the trained model.
Results on Market-1501
| Setting | mAP | Rank-1 | Rank-5 | Rank-10 |
|---|---|---|---|---|
| Standard search | 84.78% | 93.88% | 98.25% | 98.90% |
| With re-ranking | 93.69% | 94.95% | 97.65% | 98.40% |
These numbers come from the attached model_final.pth, which holds the final-epoch weights of a single 60-epoch run (seed 42), evaluated once on the test split.
Highlights
- Console commands (
reid-download,reid-train,reid-evaluate,reid-visualize) that work from any directory. - Honest evaluation, with an optional validation split and headline numbers taken from the final weights.
- Re-ranking that matches the published reference implementation.
- More than 300 automated tests, with CI on Python 3.10 to 3.14, type checking, and a Docker smoke test.
- A slim core install with optional extras for plotting, the demo, and the dataset downloader.
Using the weights
pip install "secondsight @ git+https://github.com/vardhjain/Secondsight.git@v0.2.0"
reid-evaluate --weights model_final.pth --data-root /path/to/Market-1501-v15.09.15The attached results.json, history.json, and config.yaml record the metrics, the training curve, and the exact settings of the run. See the changelog for the full list of changes and the model card for intended use and limitations.