Number of filters in Fire modules not consistent with the SqueezeSeg paper

[hagianga21]

Hi, in SqueezeSeg paper it is pointed that for an input tensor of HxWxC, the squeeze layer of the fire module has C/4 filters, and the expand layers have both C/2 filters, but in the code you are setting:
expand1x1_planes= expand3x3_planes= 4 * squeeze_planes

With this sizes, the fire module does not preserve the equality between the input size and the output size. Sorry, I am just curious :) Thanks for your help

[tano297]

Hello,
Good catch. After reading the paper I decided that the easiest way to do exactly the same network was to use their code, for a fair comparison. When I did, I did not see that discrepancy. But if you look here https://github.com/BichenWuUCB/SqueezeSeg/blob/master/src/nets/squeezeSeg.py , it looks like the author in the code is using squeeze*4 everywhere. Only some layers keep the dimensions.

I do not recommend using squeezeseg for the semantic kitti dataset. Go for one of the darknet backbones, they achieve significantly higher IoU.

Closing this, as it is not an issue. But feel free to keep commenting

[hagianga21]

Yeah, I also try the Darknet, how about the Focal loss they claim in the SqueezeSegV2 paper? Do you use it in this code?

[tano297]

I tried it, but it did not make a lot of difference vs inverse frequency scaling. That is very dataset dependent...