Fix hardcoded absolute paths

[oseiskar]

Thank you for publishing the code! The performance seems great in the benchmarks. However, to improve replicability, please consider fixing the following things:

1. Hardcoded paths to your own home / data folders:

   • SplaTAM/bash_scripts/nerfcapture2dataset.bash

     Line 4 in 7f97eeb

     python scripts/nerfcapture2dataset.py --save_path "/home/jay/4DTrack/data/iPhone_Captures/demo_1" --n_frames 1

   • SplaTAM/configs/tum/tum_eval.py

     Line 22 in 7f97eeb

     workdir=f"/storage2/datasets/nkeetha/4d/experiments/{group_name}",

   • etc

2. The offline script mentioned in the README (bash bash_scripts/nerfcapture2dataset.bash). It seems to contain parameters not supported by the Python script https://github.com/spla-tam/SplaTAM/blob/main/bash_scripts/nerfcapture2dataset.bash. In addition, it seems the depth data recorded by the current version of the Nerfcapture app is actually broken due to this issue How to use captured depth images? jc211/NeRFCapture#10 (comment). (produces large upscaled RGB PNGs instead of, e.g., 16-bit grayscale depth)

[Nik-V9]

Hi, Thanks so much for pointing this out! We have now further cleaned up the code.

All the hardcoded paths in the configs have been fixed. The bash script has also been fixed. Please try the latest code and let us know.

NeRFCapture, by default, upscales the depth to the RGB image using nearest neighbor interpolation. You can see this happening at

SplaTAM/scripts/nerfcapture2dataset.py

Line 138 in 2be31aa

sample.width, sample.height), interpolation=cv2.INTER_NEAREST)

However, we have observed that this only adds a bit of noise to the depth maps, which is fine for SplaTAM. We found the captured depth data valid, with a pixel intensity of 6553.5, corresponding to 1 meter.