Splat Trained on Normal Maps!
128
1.29MB
19w ago
A Gaussian Splat, trained on a world-space normal maps. Just for fun! Using a dataset from my upcoming video on multi-view photometric stereo reconstruction that allows you to create a laser scanner level of quality 3d mesh of a textureless object using just a smartphone on a tripod and a battery powered LED light (that and an RTX GPU, of course =)). https://x.com/pavlo_fadieiev
2 Comments
pavelf200205 19w ago
The technique for texture-less object scanning is called Multi-View Photometric Stereo. There have been quite a progress in this space recently (SuperNormal, RNb-NeUS, SDM-UniPS and Uni-MS-PS are all the new papers that i've used in my testing)! Check out this thread: https://x.com/janusch_patas/status/1896517684910805323. Note that this gaussian splat was trained just for fun (inspired by MrNeRF's post about depth maps). Gaussian splatting is not used in the actual reconstruction of the mesh. I'm planning on publishing a detailed video tutorial on how to do Multi-View Photometric Stereo.
simonbethke 19w ago
Wow, I love your hologram test :) But I didn't find any more information about the texture-less object scanning. I think this is very interesting. Maybe there is a chance to reverse the gaussian-splatting process by moving the light instead of the camera