search

Lab-of-AI-and-Robotics / GS_ICP_SLAM

[ECCV 2024] RGBD GS-ICP SLAM
MIT License
250 stars 26 forks source link

Question about rendering results. #33

Open wangm-buaa opened 1 month ago

wangm-buaa commented 1 month ago

Thank you very much for your willingness to openly share the source code! However, the rendering results i got were not like what you showed in paper. I saved the rendering rgb generated from https://github.com/Lab-of-AI-and-Robotics/GS_ICP_SLAM/blob/bb6398fd1e95b78b4197400203934989b1517cff/mp_Mapper.py#L383. The rendering results contain many holes, and the average PSNR result only has 26.47 dB (room0). Here is a example (frame_890) of the results:

final_render_00890_15 31 The final metric results are as follows: image

Riboha commented 2 weeks ago

Hello!

Sorry for the late response. GS-ICP SLAM is implemented in multiprocess manner, and optimization of the map is performed during tracking process. So the number of map optimization iteration is not constant, depending on the performance of the GPU. Therefore, the map quality can vary depending on the performance of the GPU.

The results reported in the paper are from experiments on the RTX 4090. If your environment is different from ours, your map quality might different.