List of questions about synthetic dataset preparation

[AlbertoRemus]

Hello and thanks for the amazing work!

I'm interested in reconstructing a synthetic 3D model from a set of images (like the one in the figure) and corresponding camera poses with respect to the world frame, by training your proposed model:

By following
https://github.com/NVlabs/instant-ngp/blob/master/docs/nerf_dataset_tips.md I prepared 100 200x200 images with the following "hyperparameters" (the 100 poses have been sampled from a 0.8 meters sphere pointing to the object frame)

  "aabb_scale": 1,
  "scale": 1.0,
  "camera_angle_x": 1.1495303351401578

In particular camera_angle_x = 2*arctan( image_width / 2*focal_length_x )
where focal_length_x = focal_length_x_original * ( image_width / image_width_original )

(image_width, image_height) = (200, 200)
(image_width_original, image_height_original) = (960, 540)

Here's the complete dataset for replication if needed
banana.zip

So far results are not yet satisfactory even though we can grasp the shape of the object:

Hereafter a list of questions:

• in the lego synthetic dataset only parameter camera_angle_x, why is camera_angle_ynot specified? Is it because image width and height are equal (800x800?)

• is it equivalent to use fl_x (and fl_y) parameters instead of the field of view?

• Again in the lego dataset I can see the parameter rotation for each frame, what's its meaning?

• As explained here Confusing coordinate system. #72 (comment) default scale is 0.33, so I was wondering which parameters need to be changed if the scale is set to 1.0

• I upload one last picture to show that the model seems to be properly displayed at the center of the unit-cube and to ask (a more speculative question) about the possible reasons behind the artifacts close to the camera frames