Question about GT rasterization agreement percentage

[zyuanbing]

Excellent work! Thanks for the release.
When I ran the test file after compiling the diff ras, I found the GT rasterization agreement percentage is around 60, while the masks produced by Pytorch version and CUDA version are almost same. And the whole training process works fine.

GT Rasterization agreement: 0.594480990116706
Rasterized agreement (tau = 1.0): 0.9998779296875
Gradient agreement (tau = 1.0): 1.0

BoundaryFormer/projects/BoundaryFormer/run-rasterizer-tests.py

Lines 53 to 59 in 60f580b

	segmentations = [ann["segmentation"] for ann in annotations]
	ground_truth_masks = torch.stack([rasterize_polygons_within_box([np.array(segmentation[0])], xyxy, RESOLUTION) for xyxy, segmentation in zip(xyxys, segmentations)]).float().to(DEVICE)
	ground_truth_rasterized = rasterize_instances(HARD_CUDA_RASTERIZER, segmentations, RESOLUTION)
	agreement_percentage = torch.count_nonzero(ground_truth_masks == ground_truth_rasterized) / float(ground_truth_masks.numel())
	agreement_percentages.append(agreement_percentage.item())
	print("GT Rasterization agreement: {0}".format(np.mean(agreement_percentages)))

Does this implicates there is a gap between the mask produced by HARD_CUDA_RASTERIZER and the gt mask ? And what causes this gap?
Thanks in advance.

[XJKunnn]

Hi, I think this may caused by the different values of background pixels. The value of background pixel in ground_truth_masks is 0, however it is -1 in ground_truth_rasterized.