Closed davideCremona closed 4 years ago
Hey, bgr_y should stay at True because the input images should be mapped to a canonical lighting for robustness against different lighting conditions. I think you get black reconstructions because your objects are too thin and most of the target is black. You can fix it, I quote the README:
"Middle part should show reconstructions of the input object (if all black, set higher bootstrap_ratio / auxilliary_mask in training config)"
Camera parameters are defined in the training config, you can set them arbitrarily. :)
Hi, I'm trying to make some experiments using your code but I can't explain some strange things: I've changed a little bit the code of meshrenderer.py (render function) to have different phong parameters w.r.t. the 'random_lights' flag:
Now, in dataset.py, in render_training_images() you render bgr_x and bgr_y (input of encoder-decoder network and target). The default parameters include "random_light=False" for bgr_y and "random_lights=True" for bgr_x to better generalize lights.
My experiment is to change the rendering call of bgr_y to have "random_lights=True", but when I start training the model I get these training images:![training_images_29999](https://user-images.githubusercontent.com/12844840/73475325-6d84e900-4390-11ea-9150-dbeceeff75b6.png)
Have you got some ideas to why the model is not predicting anything? The only change to the code is that flag set to True instead of False that causes different shader parameters to be considered, does this cause this problem?
I have also another question: what are the camera parameters? I need them to set a Blender camera like your OpenGL one.
Thank you in advance :)