What is the loading format of COCO training data?

I'm analyzing the COCO-format dataset loading logic in the project and I feel puzzled about the annotation format.

The standard annotation format of the original COCO dataset is (x, y, w, h), but I noticed that the code comment describes the annotation format as (xmin, ymin, xmax, ymax).
https://github.com/facebookresearch/sam3/blob/99d02f28c8d34610ab35559b72642b01f10fdb1e/sam3/train/data/coco_json_loaders.py#L18-L34

Meanwhile, the logic for calculating the bounding box area in the code is actually implemented in accordance with the (x, y, w, h) format (area = w * h).
https://github.com/facebookresearch/sam3/blob/99d02f28c8d34610ab35559b72642b01f10fdb1e/sam3/train/data/coco_json_loaders.py#L226

Looking forward to your response!

Read sam3/train/data/coco_json_loaders.py around lines 18-34 and 226, then compare the documented annotation format with the loader's area calculation and the original COCO specification. Done means resolving the discrepancy in the documentation or implementation and making the expected bounding-box format unambiguous.

Written by the indexing model from the issue text.