search

fabiotosi92 / SMD-Nets

SMD-Nets: Stereo Mixture Density Networks
MIT License
156 stars 27 forks source link

About the crop size #9

Closed yinhanxi closed 2 years ago

yinhanxi commented 2 years ago

When training,i set the crop size to 512×512 by default,but i have some problem with the pooling layer. _lib/backbone/PSMNet/submodule.py_ _output_branch1 = self.branch1(output_skip)_ _RuntimeError: Given input size: (128x32x32). Calculated output size: (128x0x0). Output size is too small_ You set a larger size when training,right?Or there is another problem. Thank you

fabiotosi92 commented 2 years ago

Hi, the crop size should be set considering the _aspectratio flag. In fact, crops are computed directly on the high-resolution images. Thus, if you want images at 512x512 as input of the network during training, you should set (for example) _cropwidth=2048 and _cropheight=2048, aspect ratio=0.25. You can find some training scripts in the script folder.

yinhanxi commented 2 years ago

thankyou for sharing such a nice work^_^

yinhanxi commented 2 years ago

another question Why Laplace and not Gauss thankyou