The instance segmentation model reported in the paper? about bdd100k

[XiaLiPKU]

Hi, the main differences arise from the changes in evaluation.

In the paper, we adopt the coco-format ground truth and use COCO scripts for evaluation.
However, since PR 36, we introduce the usage of Bitmask. Now, the evaluation runs on predictions and ground truth with the Bitmask format.

These two formats have obvious differences. For the coco format, the predictions of different instances can overlap. However, using our bitmask format, one pixel only belongs to one instance.

Currently, we have mmdet-pretrained MaskRCNNs to be published. MaskRCNN with ResNet50, trained with a 1x schedule, can achieve 16.8 in AP.

[jcdubron]

Thanks for your detailed explanation.

I adopt the coco-format ground truth for training and evaluation, which is converted from bitmask format by the official script. The Mask R-CNN with ResNet50 and 1x schedule reaches AP = 19.8. It would be great if your training code could be published.

[XiaLiPKU]

The coco-format json from polygon is different from 'the one from bitmask'. Polygons allow for overlaps, but Bitmask doesn't. So that's source of difference.

[jcdubron]

I see. Is there any script to generate coco format from polygons? Or did you convert the format with something like poly2patch function in link and then to rle with cocoapi?

[XiaLiPKU]

Yes. In scalabel, we provide this one: https://doc.scalabel.ai/label.html#to-coco.

However, we don't advise our users to do so. Because our official evaluation adopts the bitmask format.

[jcdubron]

Thanks a lot.