confusion about frame_sample_mode

[rockywind]

Hi, I am confusion about the setting of frame_sample_mode.

1. I don't know the meaning of the magical number of 4 and 2 in the code block.
2. I don't known the meaning of this for loop

               for cam in range(4):
                   ids_loss += [cam * steps + i for i in range(start_from, steps, 2)]
                   ids_render += [cam * steps + i for i in range(1 - start_from, steps, 2)]
                   start_from = 1 - start_from

3. When I train my dataset with only a front camera no stereo camera and fisheye camera, how can I modify these settings to fitting my dataset.

[Brummi]

When training on Kitti-360, we have 4 different camera views at steps timesteps. -> 4 * steps frames. The list of frames is ordered by camera. imgs = imgs_p_left + imgs_p_right + imgs_f_left + imgs_f_right (from kitti_360_dataset.py)

To make training more stable, we partition the frames of each cameras separately. This is what happens in the four loop. frame_perm[0] is a random number in [0, nv]. -> 50% chance to be even -> start_from is randomly chosen from {0, 1}.

To train on your own dataset, you can try the "default" frame sample mode, which randomly selects n_frames_render frames from the list of frames for color sampling. The rest of the frames are used to compute the loss. Just make sure that the n_frames_render value is set adequately in the config. It should be around 50% of your total frame count.