Unable to train model on custom BEV dataset

[nikhilgosala]

Hi,

I'm trying to train the provided model on a custom BEV semantic segmentation dataset, but the results are terrible. The network always collapses to the mean output and it does not seem to be learning anything. I even tried reducing the classes to contain only vehicles and road, but the results do not change. Given that my dataset has a much larger resolution (px/m) than the one used in the paper, I also tried increasing the "xbound", "ybound" and "dbound" resolution sizes with no success.

There is no issue with the dataset because I have successfully managed to train 3 other existing approaches on it.

Could you please advise as to which part of the network I should look into to debug this issue?

Thanks!

[sunnyHelen]

@nikhilgosala Hi, how did you get the BEV semantic segmentation dataset? I also wanna try to use this kind of dataset.

[VeeranjaneyuluToka]

Ok, i have been thinking on this for some time and am not sure if i posted/discussed somewhere yet. Once we design some network for a particular input resolution, can it be fed with any sized input resolution? I wonder it can not be as it changes the resptive field and might lead to bad results. Let me know if there are any different views on this.

[Camellia-hz]

@nikhilgosala Hi, I have the same problem, but my question is how to train my own data? How should I build the data.py in it? I would be very grateful if you could share your experience!