search

TRAILab / CaDDN

Categorical Depth Distribution Network for Monocular 3D Object Detection (CVPR 2021 Oral)
Apache License 2.0
359 stars 62 forks source link

How to draw the smear effect #48

Closed tonyyang1995 closed 3 years ago

tonyyang1995 commented 3 years ago

Thanks for sharing the codes. Just wonder how to draw the smear effect as Figure 1 show in the papers? image

codyreading commented 3 years ago

Hello!

Figures (b) and (c) here were generated by plotting the BEV features generated by CaDDN. That would be batch_dict["spatial_features"] stored in the batch dict. Figure (c) should be able to be plot easily, but Figure (b) was done manually setting the depth distributions to ones here rather than using the estimated values.

tonyyang1995 commented 3 years ago

Thanks, that helps a lot.

rockywind commented 3 years ago

Thanks for your help on the previous issue. I drew the smear effect image. I found that the BEV heatmap's coordination was not in conformity with the raw image. The final result was correct. I found there is exist a horizontal flip in the BEV heatmap. image

rockywind commented 3 years ago

I rotated the BEV heatmap's angle to consistency with the raw image. There exists a horizontal flip in the BEV heatmap. image

codyreading commented 3 years ago

Hi @rockywind,

Yes this is expected. This has to do with the difference between the LiDAR and camera reference frame, which are rotated and pointed in a different direction relative to one another. In order to visualize you need to perform this rotations + flip, but the transformation is still correct in the underlying code.

rockywind commented 3 years ago

Thanks, that helps a lot. @codyreading