U-net vs Tiramisu on Camvid

@jeremy I thought of using U-net on the camvid dataset, but the predictions I got weren’t very good. I trained the Tiramisu for only 2 iterations getting an accuracy of 56%. After this, I trained U-net for 30 iterations and got an accuracy of 66%.

These are the predictions I got for the same image for Tiramisu and U-net.



Despite the high accuracy of U-net, why are the predictions so bad with respect to Tiramisu?

Are you using the large tiramisu from lesson 14 ?
If so, how much memory is the model using?
I have a gtx 1070 and the large model from lesson 14 gives me an out of memory error.
I am sorry for hijacking this thread, but I am really stuck

I used a batch size of 3, so the memory doesn’t go out of bound.

What is the memory of your GPU ?

The GPU has a RAM of 12 GB. Its a Tesla K80.

uh, i am using batch size 3 as well, and it doesn’t seem to work :frowning:

I used Floydhub to run the model. If you are still getting the error try using it.

1 Like

thanks for that!
will try!

If you want, you can refer my github repo https://github.com/yashk2810/Semantic-Image-Segmentation


oh wow, thats going to be helpful, thanks

If you have any questions​, do ask!!

@jeremy Does U-net perform better when it has to segment small details like in satellite image feature detection competition on kaggle or on medical data? Is this the reason why U-net isn’t performing on the camvid dataset?

U-net and tiramisu should both be particularly good at that - see how they perform well on trees, for instance.

I don’t see why U-net wouldn’t go well on Camvid - although it might just be that there’s not enough data for it train well. Densenet (tiramisu) is good at handling small datasets.

UNet is pretty good at small datasets as well.

@yashkatariya - Try varying the block depth (number of layers), as well as the network depth (number of unet blocks) to see if that has any effect.

An experiment I’ve been wanting to get around to was using densenet style linear filter growth on unet instead of 2^x to see what effect it has generally. With all the new papers coming out on how our networks’ parameter space are potentially doing memorization, I find myself drawn to simpler and smaller networks like ENet again.

Btw, your code is very nicely organized.

Btw, your code is very nicely organized.


@haresenpai I increased the network depth of U-net and it was able to improve the predictions by a small margin.

Code link:- https://gist.github.com/yashk2810/bd317f854794d2d6969240420bb1407d. The predicted image is at the bottom.

@jeremy The picture had a tree, but it wasn’t able to segment it.

I guess by increasing the network depth and number of layers per block as @haresenpai said, the predictions will improve. But we would need to add many layers for this to work.

@jeremy Is my conclusion right?