Test on my data

[YihuanL]

Hello,
I have run your code in Matlab, which shows good performance with your test data(samples_ITOP_frontal_noseg.mat). Here, I want to know how to save my depth image(*.png format) to this *.mat format so that I can test with it?
thanks.

Hello, I could run the code with my data using the code below to generate a .mat file for neural network input. But the predicted skeleton on my data are erroneous... maybe it is due to my image resolution which is a lot higher than the one of ITOP images (the resize to 100x100 is maybe too hard).

clear all
clc

%Get input depth map data single format expressed in meters then rescale to
%100x100 resolution and [0,1] values

dir_path='yourpath/';
depth_files_path=fullfile(dir_path, '*.png');
depth_files=dir(depth_files_path);
numfiles=length(depth_files);
addpath('natsortfiles'); %download natsortfiles for matlab on https://nl.mathworks.com/matlabcentral/fileexchange/47434-natural-order-filename-sort
depth_files_sorted=natsortfiles({depth_files.name});

for k=1:1:numfiles

depth_map_path=fullfile(dir_path, depth_files_sorted{k});
depth_map=imread(depth_map_path);
depth_map=single(depth_map);
depth_map=depth_map./1000;
 
depth_map=imresize(depth_map, [100, 100],'nearest');
addpath('Utils/');
depth_map=mj_rescale(depth_map, 0, 1);
samples.data(:,:,k)=depth_map;

end

%Add the pose dictionary for the algorithm learned on ITOP

itop_data=load('yourpath_to_deepdepthpose/ddp-master/Data/samples_ITOP_frontal_noseg.mat');
samples.centroids=itop_data.samples.centroids;

%Add joints ground-truth locations (here I put a zero matrix since I do not have annotation I am just doing visual inspection of pose estimation results)

joints_groundtruth=zeros(45,numfiles,'single');
samples.joints_gt=joints_groundtruth;
save('Data/yourinputdata.mat','samples');

[YihuanL]

Thanks for your reply, I will test that.

Hello @YihuanL,

Did you manage to use Deep Depth Pose on your own data and get good pose estimation with it?

Thanks a lot :)

[YihuanL]

hi @Miaou446 ,
Yes, I have tested my own data with your help, the result (estimated pose) is also incorrect. I think maybe it is the centroids of samples caused this problem, because you extracted them from 'samples_ITOP_frontal_noseg.mat' by using this code:
samples.centroids=itop_data.samples.centroids
While, I have no ideas about how to get the centroids .

Hi, @YihuanL,

The centroids is the dictionary of poses learned by clustering on ITOP dataset.
If we take similar poses as the ones in ITOP in our own dataset it should though work. But it seems it is not the case.

I have also no ideas about how to get the centroids in same format used by deep depth pose for our own data.

[YihuanL]

Hi, @Miaou446 ,
If you just want to find a good method to extract skeleton from depth image, the code in following link shows a good performance when I test it with my own data.
https://github.com/mks0601/V2V-PoseNet_RELEASE

Okay, thank you!

[otepencelik]

Hi, @Miaou446 ,
If you just want to find a good method to extract skeleton from depth image, the code in following link shows a good performance when I test it with my own data.
https://github.com/mks0601/V2V-PoseNet_RELEASE

Hi @YihuanL . How did you manage to test your own data with the link that you provided (PoseNet)? I don't have any experience with Lua programming so I couldn't understand their repo at all. Can you share with me some sample codes?