search

spla-tam / SplaTAM

SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM (CVPR 2024)
https://spla-tam.github.io/
BSD 3-Clause "New" or "Revised" License
1.59k stars 174 forks source link

Custom iPhone Data Performance #65

Closed wangjinhoon closed 8 months ago

wangjinhoon commented 10 months ago

The image below shows the result generated from 500 pieces of data created with an iPhone. Despite inputting gt_data with poses, the results are good for the wall but not for objects inside the room like desks. I'm curious about your thoughts on why this might be the case. image

Nik-V9 commented 9 months ago

Hi, there could be many reasons. My initial guess is the noise and incorrect maximum depth from the iPhone. Could you share this data with us?

Also, I'll shortly be pushing code for Gaussian Splatting with iPhone data. That should provide further insight into why things are failing.

wangjinhoon commented 9 months ago

Hello, thank you for responding. Below is the data I used. Could there have been an issue with the way I generated my data?

https://drive.google.com/file/d/1qjMWTfspVbYGDg-pZ9NDDw_WerZRyxAF/view?usp=sharing

and Could you please let me know how to see Progressively Growing of Gaussians in real-time and what parts I need to modify?

Nik-V9 commented 8 months ago

Hi, I tried doing large-scale reconstructions similar to your shared data. I've observed that the iPhone data has a lot of artifacts and incorrect depth for scenes with a scale greater than 5 meters.

We will have a follow-up fixing the reliance of SplaTAM on good depth.

wangjinhoon commented 8 months ago

Thank you for your concern!