From my own dataset, I can't replicate your results, what should I think about this?

[sunshineywz123]

[sunshineywz123]

How is our depth map generated?

[bennyguo]

Hi! What dataset are you using? If it's in blender format, could you please share the transforms.json file? Thanks!

[sunshineywz123]

Hi! What dataset are you using? If it's in blender format, could you please share the transforms.json file? Thanks!

blender

[sunshineywz123]

Hi! What dataset are you using? If it's in blender format, could you please share the transforms.json file? Thanks!

transforms.zip

[bennyguo]

Could you try to pull the latest code and replace the following lines https://github.com/bennyguo/instant-nsr-pl/blob/4f70db328827dee6596f1553df0177255f32a1c2/datasets/blender.py#L39-L43 with

self.directions = \
    get_ray_directions(self.w, self.h, meta['fl_x']), meta['fl_y']), meta['cx']), meta['cy']), self.config.use_pixel_centers).to(self.rank)

and adapt img_wh in the config file to your setting (for example 960x720)?

[sunshineywz123]

Could you try to pull the latest code and replace the following lines

https://github.com/bennyguo/instant-nsr-pl/blob/4f70db328827dee6596f1553df0177255f32a1c2/datasets/blender.py#L39-L43

with

self.directions = \
    get_ray_directions(self.w, self.h, meta['fl_x']), meta['fl_y']), meta['cx']), meta['cy']), self.config.use_pixel_centers).to(self.rank)

and adapt img_wh in the config file to your setting (for example 960x720)? Ok, I'll try, what is causing my current problem and how to understand this depth map?

[bennyguo]

I'm not quite sure. If you set img_wh to be 800x800 in your experiment, there could be a mismatch between the camera intrinsics and the image resolution, leading to wrong rays_o and rays_d.

In the depth map, blue means near and red means far.