shubhtuls / mvcsnp Goto Github PK

Hi,
when following the demo file to visualize the meshes in https://github.com/shubhtuls/mvcSnP/blob/master/renderer/meshUtils.lua#L206 it tries to open render.sh, which doesn't exist. Maybe renderPose.py is meant? I tried replacing it, but it didn't help.
The error is bash: ../renderer/render.sh: No such file or directory

Hi, @shubhtuls , I am very interested in your work. When I read your code named shapenetPoseReg.lua , I can't understand encoderPose network. Therefore, I print it.

Could you tell me the intention of (23): nn.gModule ?Why is it so complicated?

Hi @shubhtuls , thanks for the code. But I have some problems generating the dataset:

1. I am using Python3.6 instead of Python2.7. So I used _pickle instead of cPickle in renderPreprocessShapenet.py. But I got error in Line46 of this file: "TypeError: can't pickle dict_key objects". I found that this line might be ignored, so I comment it and keep running the code.

2. I also found that in Line263 of renderer/render_model_views.py, rl.outputs doesn't have the key 'Z', I replace it with 'Depth'. I wonder if this is the version problem. I used the latest 2.79 version of blender. And I download the shapnet models using http://shapenet.cs.stanford.edu/shapenet/obj-zip/ShapeNetCore.v1/<synsetId>.zip (I only downloaded the 3 categories used in the paper).

3. Later I found that my render_x.png are all black pixels. I then check the .exr file, the rgb channels are all 0 values. Not sure what's the problem.

4. I am not familiar with blender, I wonder if it can also generate surface normal output.

Thanks!

When testing with ground truth pose and ground truth shape, this loss is not zero, so can you explain the reason for me? Thx!

Hi @shubhtuls , seems that the default projection type is perspective, I wonder what should I modify so that I can get the calibration matrix of a orthographic projection camera.

Hi,
Me again from discussion about your "drc" paper.
Since predicted frames can be arbitrary, I wonder what we can do to make them align consistently. In your paper, the event cost for depth image is difference between depth associated with sample point and real depth. In this formula, if we replace depth associated with sample point, i.e., instead of using distance to sample point, we use distance to the voxel where the sample point settles.

Our hope is to minimize distance to voxel rather than sample point. If we use this replacement, can predicted frame be aligned consistently? What's your opinion? Looking forward to your reply

When I execute preprscessing.md , the render_.png and depth_.png can't be generated normally. Then I debug BlenderRenderer.py line 93, like this
(a, b, g, r, dmap) = [fromstr(s, width, height) for s in exrimage.channels('ABGRZ', pt)]
the (a, b, g, r) are all zero, the dmap's max is 100000000000.0 and min is 1.7825699.
The problem is why (a, b, g, r) are zero?