erogol / seg-torch

Segmentation with deep learning

Seg-Torch for Image Segmentation with Torch

This work was sparked by my personal research on simple segmentation methods based on deep learning. It is the harvest of two great predecessors;

• https://github.com/e-lab/ENet-training
• https://github.com/fedor-chervinskii/segnet-torch
  • (real project page) http://mi.eng.cam.ac.uk/projects/segnet/

However this code includes radical differences (such as data loading, augmentation, memory optimization) and it has more generic type of implementation suitable for use in any custom project. You only need to modify data-loader files data/custom-gen.lua and data/custom.lua.

Be warned this is susceptible to bugs. Any pull request is appreciated.

Check train_scripts/ for example execution.

Models

• SegNet: Very simple encoder-decoder network, segmenting end2end
• EroNet: Very similar but it chops Batch-Normalization and uses ELU activation. It is lower in accuracy but faster in training.

Example Results

exp_model/ includes a proof of concept on CamVid dataset. If you compare the results with the real-project this implementation has higher values interestingly (at least for me) :) .

Model will be shared on Dropbox, as soon as I find some time to do so.

seg-torch's Issues

bugs in the code

Thanks for sharing the excellent work. In camvid.lua file, I think it is better to remove self.dir or set it to be nil.

seg-torch/data/camvid.lua

Line 14 in d4ce812

self.dir = opt.datapath

because it will give us the wrong directory where the image data sits., which will be:

self.dir/self.dir/sub_path_to_images. 

according to the line in

seg-torch/data/custom-gen.lua

Line 33 in d4ce812

table.insert(imagePaths, line)

variable line(which represents path) will contain self.dir as its prefix substring

Segnet output/target dimensions

Hi,
thank you for such a great work!
Could you please explain the required dimensionality of outputs and targets? Segnet model input is [batchSize, numChannels, H, W] and the output is [batchSize, numClasses, H, W]. But cudnn.SpatialCrossEntropyCriterion() requires the target to be of size [batchSize, H, W]. So it looks like that the model outputs numClasses separate masks for each class while the criterion expects a single mask with all the classes within it.
I can't get through this issue while reproducing the model with a toy example.