image-depth misalignment question

Hi,

In the paper, you identify several types of ground-truth issues in the dataset, such as image–depth misalignment, invalid backgrounds (e.g., skyboxes), clipping artifacts, and others. However, the detection method appears somewhat heuristic. For example, to filter samples with image–depth misalignment, you apply Canny edge detectors to both the grayscale image and the depth map to extract boundaries. The boundaries are first dilated by 1 pixel, and the percentage of intersection between the image and depth edges is computed. The boundaries are then further dilated by 3 pixels, and the intersection is computed again. Finally, the ratio between the two intersection scores is used, and samples whose ratio falls below a certain threshold are removed.

Could you provide more justification for choosing this specific method and parameter setting? In particular, it would be helpful to understand why this metric is effective for detecting misalignment. Additionally, have you considered or compared alternative approaches (e.g., direct edge alignment metrics, chamfer distance between edge sets, or cross-correlation-based alignment measures)? Including such comparisons or ablation studies would help clarify the advantages of the proposed heuristic.

Thanks a lot!