The pretrained KITTI-2015 model

[youmi-zym]

Hi, thanks for sharing your code. I have some questions need your help. my environment: pytorch 0.4.0 torchvision 0.2.0

I just downloaded your pretrained params and cloned your code, then executed this command python submission.py --datapath [kitti_scene_flow_path]/testing/ --loadmodel pretrained_model_KITTI2015.tar without any fine-tuning. Then, I used the disp_read.m and disp_write.m in matlab code to rewrite the result. Finally, I submit to the KITTI website to evaluate. Blow is the result,

Mine:
All / All   2.36    5.72    2.92
All / Est   2.36    5.72    2.92
Noc / All   2.19    5.42    2.72
Noc / Est   2.19    5.42    2.72

Yours:
All / All   1.86    4.62    2.32
All / Est   1.86    4.62    2.32
Noc / All   1.71    4.31    2.14
Noc / Est   1.71    4.31    2.14

You can find that there is 0.6% difference between us. From the result picture, it almost similar, but my error map is shallow than you in color.

I also executed this command python submission.py --datapath [kitti_scene_flow_path]/training/ --loadmodel pretrained_model_KITTI2015.tar and then used the matlab code to evalute the result, I recorded every image's disparity error in exp.txt file. the average disparity error is 0.95%. And I also gave one of them blow. The error map is shallow as the testing result. I think there must be some mistakes made, please check, thanks a lot! exp.txt

[IceTTTb]

Hi, @youmi-zym It seems that there is something wrong to your error map. These are two my results on KITTI2015 val set.

[youmi-zym]

@IceTTTb Thanks for your advice. To be actually, I use the matlab toolkit from the KITTI website. and blow is the main part of the code I have rewritten.

for i = train_val_list
    D_est=disp_read(fullfile(train_root, iids(i).name));
    D_gt=disp_read(fullfile(gt_root, iids(i).name));
    rgb=imread(fullfile(img_root, iids(i).name));
    rgb=double(rgb)/255.0;
    d_err = disp_error(D_gt,D_est,tau);
    D_err = disp_error_image(D_gt,D_est,tau);
    D_est_color=disp_to_color(D_est);
    D_gt_color=disp_to_color(D_gt);
    save_path=fullfile(train_root,['../error/',iids(i).name]);
    imwrite([rgb;disp_to_color([D_est;D_gt]);D_err],save_path);
end

As you can see, disp_to_color([D_est;D_gt]), the max disp to scale the color map is calculated from ground-truth and estimated disparity. Therefore, the processions I have taken are:

1. download pretrained_model_KITTI2015.tar
2. execuate submission.py, get the testing and training data result.
3. use the matlab code from KITTI website, disp_read.m -> disp_write.m to rewrite the result
4. use the matlab code from KITTI website, demo.m(show above) to get the dispairty map, there is the result of one of example you give

I'm wondering are there any errors in submission.py or environment I set or some steps I missed.

[IceTTTb]

@youmi-zym Hi, I guss there may be some errors in the third step in your processions. I feed the results from submission.py to demo.m directly, and the error maps seem right.

[youmi-zym]

@IceTTTb It meas that you got the similar map as I show above? Well, in case there are some mistakes I forgot, I will list the code and environment here:

pytorch 0.4.0 torchvision 0.2.0 python 3.5

disp_read.m

function D = disp_read (filename)
% loads disparity map D from png file
% for details see readme.txt

I = imread(filename);
D = double(I)/256;
D(I==0) = -1;

disp_write.m

function disp_write (D,filename)
% saves disparity map D to png file
% for details see readme.txt

D = double(D);

I = D*256;
I(D==0) = 1;
I(I<0) = 0;
I(I>65535) = 0;
I = uint16(I);
imwrite(I,filename);

[passion3394]

I have rewritten the disp_to_color.m in Python, welcome issues for somebody else, I will try to make the script running fast.

https://github.com/passion3394/PSMNet_CPU/blob/master/disp_to_color.py