DetectionDataset.from_coco → as_coco → from_coco was not idempotent. The == comparison after a save/reload cycle returned False because detections_to_coco_annotations discarded annotation metadata from detections.data.

Root causes in detections_to_coco_annotations

• area always overwritten with bbox area — box_width * box_height was used unconditionally, ignoring detections.data["area"] which holds the original COCO annotation area (typically polygon area, not bbox area).
• iscrowd stored as Python bool — contains_holes(...) or contains_multiple_segments(...) returns a bool; COCO expects int (0/1). Serialized as true/false in JSON.
• iscrowd reset to 0 when no mask — crowd annotations (iscrowd=1) loaded from COCO were silently demoted to 0 on save if no mask was present.

Fix

# Before
iscrowd = 0
if mask is not None:
    iscrowd = contains_holes(mask=mask) or contains_multiple_segments(mask=mask)
...
"area": box_width * box_height,
"iscrowd": iscrowd,  # Python bool or 0

# After
if mask is not None:
    iscrowd = int(contains_holes(mask=mask) or contains_multiple_segments(mask=mask))
    ...
else:
    iscrowd = int(data.get("iscrowd", 0))  # preserves crowd annotations

area: float = float(data.get("area", box_width * box_height))  # preserves original area

When detections.data has no area/iscrowd (e.g., detections not loaded from COCO), behavior is unchanged — falls back to bbox area and 0 respectively.

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.