- Resolve issue where
CropProjector().get_embeddings()failed occasionally when intermediate batches of encoder had single instances.
- Revamp Overview section. Collate error counts into colored bars.
- Use HDBSCAN instead of DBSCAN to cluster crop visualizations.
- Group visualizations within clusters by sub-clusters.
- Combine Classification Errors and Classification And Localization Errors into a single section (Confusions).
- Update bounding box size visualization in Missed Errors section.
- Use a violin plot instead of a box plot.
- Use a logarithmic scale for the y-axis.
- Add jitter to the data points.
- Group missed errors by likely causes.
- Standardize colors of bounding boxes for each error type.
- Define new comparison report to compare the performance of two models.
- Defined new module (
gt-flow) to visualize changes in ground truth predictions between modules. - Order image samples by visual similarity, copmuted using DBSCAN. Identify commonalities between failing cases more easily.