happinesslz / EPNet

EPNet: Enhancing Point Features with Image Semantics for 3D Object Detection(ECCV 2020)
MIT License
235 stars 36 forks source link

about data augmentation #35

Open pummi823 opened 2 years ago

pummi823 commented 2 years ago

Dear author: you mention that you use three data augmentation strategies include: rotate, flip, and scale transfer the point clouds. But, in your code, you look like have not do the same augmentation for the input image, will there occur wrong when you fusion the pc with img?

Rtakaha commented 1 year ago

Thank you very much for sharing your work, @happinesslz !

I have the same question. Data augmentation is applied here, and AUG_DATA is True when the given config is used.

Rotation and scaling might not cause big misalignments if the parameters are small enough, but flipping should cause strong misalignments.

How should I align img and pcd? Thank you.

Rtakaha commented 1 year ago

Sorry that I didn't check previous issues enough. Now I understand how img and pcd are aligned. As this comment mentions, pcd and img are aligned here, by storing the xy coordinate on image plane of each point.

Thank you!