Reproduce the result of localnet

[s7ev3n]

Hi,
Thanks and waiting for your code release.

I am trying to reproduce the result of localnet. The localnet takes 1216x256x1 velodyne_raw(not normalized,but divide 256 to obtain real depth value) as input, and a simple conv layer with stride=2, kernel_size=3, out_channels=32, following two hourglasses as in the paper, then a tranconv layer of stride=2, kernel_size=3, out_channels=32 and a 1x1 conv layer with out_channels=2, the confidence map after softmax adds on the local depth prediction.
I trained on kitti depth dataset with adam and lr=0.001 for 40 epochs and 48 batchsize, the loss is masked mae loss. However, my result is much worse, the best rmse is around 1800 compared with your localnet result 995.

Is there any insight to improve the localnet? Thanks so much :)

-s7ev3n

[Wouter Van Gansbeke]

Hi @s7ev3n,

Thank you for waiting. I updated the code in this repo. Feel free to take a look. It's easier to spot the differences this way.

Kind regards,
Wouter