Object Detection

[Hendrik-code]

Add support for bounding boxes

• BBOX class with multiple instantiors
• Transfer everything to one default internal representation (pixel-wise, not relative!)
• Internally track what the data originally was (bbox or mask), but be able to convert back and forth
• Add all corresponding metrics (AP, mAP, RoDeO, ...).
• Add warnings if you calculate seg metrics for bbox representation and vice versa
• Hungarian Matching
• Metrics: Localization score (gaussian distance to COM, could be used for seg as well)
• Shape score (overlap score when centering both by COM)
• Coco style rather bad, torchmetrics also not flexible enough
• Support for Object Score Handling (AUTC over this threshold!)
• Avoid pitfalls (Accuracy per TP, ...)
• Requires #154

Material:

• https://raw.githubusercontent.com/rafaelpadilla/Object-Detection-Metrics/master/paper_survey_on_performance_metrics_for_object_detection_algorithms.pdf https://github.com/rafaelpadilla/Object-Detection-Metrics https://github.com/rafaelpadilla/review_object_detection_metrics

[neuronflow]

cocoMAP: https://cocodataset.org/#detection-eval