search

zaiweizhang / H3DNet

MIT License
211 stars 25 forks source link

Question about data augmentation on ScanNet #19

Closed Na-Z closed 3 years ago

Na-Z commented 3 years ago

Hi zaiwei,

When performing data augmentation on ScanNet, you also flipped the height. https://github.com/zaiweizhang/H3DNet/blob/e89d092bdf4c2ab1df576e6c4ff12bcbc2d55b2e/scannet/scannet_detection_dataset_hd.py#L188 https://github.com/zaiweizhang/H3DNet/blob/e89d092bdf4c2ab1df576e6c4ff12bcbc2d55b2e/scannet/scannet_detection_dataset_hd.py#L195 May I know the reason? As far as I know, the height should not be flipped.

Additionally, you set the heading angle as the rot_angle (since original heading angle is 0 in ScanNet) https://github.com/zaiweizhang/H3DNet/blob/e89d092bdf4c2ab1df576e6c4ff12bcbc2d55b2e/scannet/scannet_detection_dataset_hd.py#L202 However, the rotate_aligned_boxes function returns params of axis-aligned bboxes. I don't understand why you set heading angle to rot_angle instead of 0 here. Could you explain?

Thank you.

zaiweizhang commented 3 years ago

Hi Na-Z,

meta_vertices[:, 6] = -1 * meta_vertices[:, 6] This line is actually flipping the angle. And this line: (meta_vertices[:, 6] += rot_angle ) is applying a global rotation so it would not affect individual objects' front orientation prediction.