Confidence scores

Each OBB can carry a score in [0, 1] for triaging a converted dataset: high scores are tight SAM fits, low scores are worth a look, and 0.0 marks a fallback where the HBB was kept. The score is a heuristic for ranking boxes, not a calibrated probability.

How it is computed

The conversion score is the product of two factors:

• Rectangularity: the area of the contour the box was fitted to, divided by the area of the box. 1.0 for a perfectly rectangular object; lower when the mask is ragged or the box encloses much empty space.
• Ensemble consensus: the area of the voted mask divided by the area of the union of the models' masks. 1.0 for a single model; for two models it equals their mask IoU.

Where fragments were fitted together, rectangularity uses the pieces' own area, so the gap between them lowers the score.

Writing it

Flag	Where the score goes
--save_confidence	a 10th column on each OBB line, and a trailing column on each polygon line
--confidence_dir DIR / -cd DIR	DIR/<image stem>.txt, one score per line, row-aligned with the labels
--confidence_dir with no value	the same, in <img_source>/../labels_confidence

Use --confidence_dir when the label files have to stay standard: Ultralytics and other YOLO OBB readers reject a 10th column. The two flags can be combined.

hbb2obb-view and hbb2obb-convert read the scores from either place. In --save_img overlays the OBBs are always tinted green to red by score, and --show_confidence also prints the value.

Choosing which score is reported

When the HBB files carry a detector confidence in a 6th column, --confidence_source / -cs chooses what the reported score means:

Value	Reported score
conversion (default)	the conversion score above
detector	the detector confidence read from the HBB input
combined	the product of the two

The conversion score says how well the box fits the segmented shape; the detector score how sure the detector was that there is an object at all. A box whose input line has no 6th column reports its conversion score, with a warning.

hbb2obb-detect writes that 6th column for the boxes it finds and, with --merge_with, for boxes drawn by hand. The sample data ships detector output, and its labels_obb/ is scored combined.

Using it

• Review the lowest scores first. hbb2obb-view images --show_confidence colors every OBB by score; c toggles the coloring in the window.
• Flag boxes for a release. hbb2obb-convert ... --to dota --difficult_from confidence --difficult_below 0.5 sets the DOTA difficult flag on every box below the threshold. See The difficult flag.
• Find fallbacks. Every run reports how many there were; a 0.0 score in the files marks which, unless --confidence_source is detector.

In Python

hbb2obb(..., return_confidence=True) returns the scores beside the OBBs, and confidence_source selects them as the flag does. See hbb2obb().