ICNet for Real-Time Semantic Segmentation on High-Resolution Images, ECCV2018

any way I can use this without having the beloved MATLAB? I love so much closed source that I am really close to buy a licence but hey ... no actually.

Hi, Hengshuang! This is a valid work for practical application. I've run your code. But something may be wrong. The predictions of network are all road, i.e. every element of 'imPred' is one. I can not locate the problem. I tested many images and got same results. What's wrong with that? thx

Hi @hszhao , I am evaluating the ICNet, and run eval_all.m on the evaluation/samplelist/cityscapes_val.txt.

Model using to evaluate:

model_weights = 'model/icnet_cityscapes_train_30k.caffemodel'; %trainval_90k for testset
model_deploy = 'prototxt/icnet_cityscapes.prototxt';

Here is my result on ICNet train on trainset for 30K, evaluated on valset (mIoU/pAcc):

==== Summary IoU ====
  1             road: 0.0000
  2         sidewalk: 0.0045
  3         building: 0.0109
  4             wall: 0.0025
  5            fence: 0.0438
  6             pole: 0.0101
  7    traffic light: 0.0000
  8     traffic sign: 0.0001
  9       vegetation: 0.0008
 10          terrain: 0.0047
 11              sky: 0.0000
 12           person: 0.0031
 13            rider: 0.0008
 14              car: 0.0012
 15            truck: 0.0000
 16              bus: 0.0000
 17            train: 0.0000
 18       motorcycle: 0.0005
 19          bicycle: 0.0000
 
Mean IoU over 19 classes: 0.44%
Pixel-wise Accuracy: 0.51%

Is there anything I am missing to run this model, thanks in advance for anything input.

By the way, does anybody be able to reproduce the result, and if yes, please let me know, what was missing?

Thanks Huaqi

Could you provide train architecture? I want to implement your idea at TensorFlow, but I cannot get your idea of network only using information from paper. There are lacking information like:

1. Number of filters in each layer, especially in pipeline image/4/
2. I released prototxt I do not see any image/4 as input, where it go?
3. I see that you are using residuals, are they the same like in PSPNet?

I was using this visualization as base, but I'm not able to match this architecture with that from pipeline.

If such error occur, try: $ ldd /ICNet/PSPNet/matlab/+caffe/private/caffe_.mexa64 and in matlab:

! ldd /ICNet/PSPNet/matlab/+caffe/private/caffe_.mexa64

You will find libQtOpenGL, libQtGui, libQtCore is point to /bin/glnxa64 but libQtTest is point to sth in /lib64/

However, all 4 lib should use the one in /lib64/

solution: delete libQtOpenGL, libQtGui, libQtCore in /bin/glnxa64 create symbolic link of these 3 and point to the one in /lib64/

Modify the related paths in 'eval_all.m':

Mainly variables 'data_root' and 'eval_list', and your image list for evaluation should be similar to that in folder 'evaluation/samplelist' if you use this evaluation code structure.

How I modify the root_path and eval_list??

I have dataset in my laptop test train and val so how I Put my dataset path on data_root?

Please guide me what is the issue actually

Hi As per step 3. evaluation

downloaded cityscapes dataset from https://drive.google.com/file/d/1z_2yifjYICczhZjBqCEhE1m6GYwJuljn/view

While running the ./run.sh, I got the below error

Error using importdata Unable to open file.

Error in eval_sub (line 3) list = importdata(fullfile(data_root,eval_list));

Error in eval_all (line 35) eval_sub(data_name,data_root,eval_list,model_weights,model_deploy,fea_cha,crop_size_h,crop_size_w,data_class,data_colormap, ...

Please help

hello, I want to train ICNet with pascal voc dataset in caffe. I have changed the number of class in .prototxt, and written the solver.pototxt according to the paper. But the loss also around 0.5, and the accuracy is too low. Because my GPU is 8G memory, I can only set mini_batchsize to 8, but by setting iter_size to 2. So the batch size is 16. Can you help me. Looking forward to your reply！

How do you set the weights for different classes when calculating loss? I train the model using the weights I give. But the model still learn the detect the background. Can you introduce the method you use?

Hi, I once used this script to test the speed of ICNet in Matlab :

tic; score = net.forward({input_data}); temp = toc;

But I got the result of 15.6fps on 1080ti without other programs running, the result provided by the author is more than 30fps, then I used the scripts provided by the author in Caffe, the result is 31fps on 1080ti, which is lager than the original result, I was confused about it. Can you help me?