search

Megvii-BaseDetection / BEVStereo

Official code for BEVStereo
MIT License
252 stars 14 forks source link

different image sizes #15

Open NorthSummer opened 1 year ago

NorthSummer commented 1 year ago

Hi, thanks for opening access. By the way, where should I modify if I want to train with 1408x512 images? I only modified the "final_dims" and the results seem to be not right. Does the depth_gt need to be generated again when changing the input image size?

yukaizhou commented 1 year ago

Have you solved this problem?

NorthSummer commented 1 year ago

Not yet, I trained BEVDepth official code with 1408x512 resolution a few days ago, but the results seemed like overfit (14.2 total loss, 29.4 NDS). However it's not reasonable...

yukaizhou commented 1 year ago

I also encountered a similar problem. Only the size of the input image was modified for the official code, without any other changes, and the result was very abnormal. Regarding what you said about regenerating the point cloud ground truth, I confirm that the code and image size are irrelevant. The loss value of my training eventually oscillates between 8.2 and 9.2. Do you think this is normal?