How to test the pixel-nerf on our own scene

[Xianjin111]

Thank you for your perfect work. I want to apply the pixel-nerf on our scene which contains the extrinsic and intrinsic. Could you give me some clues how to replace the input by our multi-view data and calibration data and which pre-trained model would be suit for scenes which are different from chair and car.

Thank you very much.

[nishant34]

Do you know now how to do it?

[sxyu]

Depends on the case. I'd say the most relevant pre-trained model may be the cross-category one. Unfortunately this was using 64x64 images and rather old renderings.

To load your own data, please adapt one of the data loaders, or perhaps format your data in DVR's format.. Store files in <object_name>/image/xxxx.png (number 0-padded to 4 digits) and camera info in cameras.npz. This should have 4x4 camera-to-world pose matrix 'world_matinv' and 4x4 'cameramat' intrinsics matrix for each view i (starting from 0). You can download the NMR dataset to see an example of the format