search

dvlab-research / UVTR

Unifying Voxel-based Representation with Transformer for 3D Object Detection (NeurIPS 2022)
227 stars 18 forks source link

Inquiry about 'Effect of Height in Voxel Space' in Sec. 4.2. #10

Closed zjwzcx closed 2 years ago

zjwzcx commented 2 years ago

Hi Yanwei, thanks for your awesome work! @yanwei-li

You've showed the effect of height in Table 1, which demonstrated larger height values along axis Z contribute more for camera-based 3d detector. I'm very curious about the performance of camera-based detectors when we set height = 21, 41 or even larger. Have you conducted some experiments with much larger height? Do you have any insight about the setting of height value?

Look forward to your reply!

yanwei-li commented 2 years ago

Hi, that's a good question. I did not try such a large height due to the GPU memory constraint. Because the height 11 is actually the same resolution of the voxel grid as that in BEV (~0.2m). I guess the performance could be saturated with a much larger height.

zjwzcx commented 2 years ago

Hi, that's a good question. I did not try such a large height due to the GPU memory constraint. Because the height 11 is actually the same resolution of the voxel grid as that in BEV (~0.2m). I guess the performance could be saturated with a much larger height.

Got it! Thanks for your reply!