Rendering results on Carla Dataset

[shairatabassum]

Hi, thank you for this wonderful work!
I'm currently trying to train mars on Carla-captured dataset by converting them to VKITTI2 format and using the mars-vkitti-car-depth-recon method. The results are not good and a total mess in most cases.

A very limited number of frames look like this:

While most of them are as follows:

Could you please advise what could have possibly gone wrong? I'm expecting the camera parameters to be quite accurate as they are retrieved directly from Carla. There are a total of 700 images in the training set with two cameras and no moving vehicles on the scene.

Also, does the training dataset size have to do anything with it, as reducing number of training images to only 40 frames makes the rendering better.
(first image with all frames with over 400K iterations, second image with only 40 frames after 40K iterations)

I would really appreciate your feedback, thank you!