search

astra-vision / MonoScene

[CVPR 2022] "MonoScene: Monocular 3D Semantic Scene Completion": 3D Semantic Occupancy Prediction from a single image
https://astra-vision.github.io/MonoScene/
Apache License 2.0
708 stars 69 forks source link

vox2pix #65

Closed yty-sky closed 1 year ago

yty-sky commented 1 year ago

The vox2pix code is that: projected_pix, fov_mask, pix_z = vox2pix( T_velo_2_cam, cam_k, self.vox_origin, self.voxel_size * scale_3d, self.img_W, self.img_H, self.scene_size, )
This operation will project the voxel to the 0th camera, not 2nd camera, Is there a mistake?

anhquancao commented 1 year ago

Hi @yty-sky, you are correct. It is a mistake. I tried to change to the 2nd camera and retrained but nothing change. I think it is due to the fact that the two cameras are parallel and very close two each other. Also, we unproject the extracted feature with large receptive field, thus it also cover the correct pixels.

yty-sky commented 1 year ago

Hi @yty-sky, you are correct. It is a mistake. I tried to change to the 2nd camera and retrained but nothing change. I think it is due to the fact that the two cameras are parallel and very close two each other. Also, we unproject the extracted feature with large receptive field, thus it also cover the correct pixels.

Yes, it is. Thank you for your reply