search

yanivw12 / gs2mesh

[ECCV 2024] Official implementation of the paper "GS2Mesh: Surface Reconstruction from Gaussian Splatting via Novel Stereo Views"
130 stars 6 forks source link

Poor performance in custom nerf synthetic dataset #15

Open Jackson334-prog opened 4 days ago

Jackson334-prog commented 4 days ago

Hi, Nice work! Your work is amazing. Most of datasets was fine. And I followed your advice and tried testing the NeRF dataset with an initial poses, setting TSDF_scale to 0.1. Unfortunately, I ran into some issues with holes appearing in the results. I also tried tweaking the TSDF_baseline, but it didn’t seem to help. image Here’s the code I’m running. By the way, I do some change in --pose_konwn arg like your doing with mobiletrick dataset python run_single.py --colmap_name zoo --skip_video_extraction --renderer_baseline_percentage 12 --TSDF_min_depth_baselines 3 --TSDF_max_depth_baselines 11 --pose_known --TSDF_scale 0.1 and the dataset download link is in here: https://drive.google.com/drive/folders/1dK2rtUwSK8RtiBcgRsTJvsuc_X3TjlqL?usp=drive_link Any response would be appreatiate!

yanivw12 commented 3 days ago

Can you please send the COLMAP sparse model after you've run it? With the full cameras.txt, images.txt and points3D.txt?

Jackson334-prog commented 3 days ago

Hi, I 've sent whole sparse in https://drive.google.com/drive/folders/1dK2rtUwSK8RtiBcgRsTJvsuc_X3TjlqL?usp=drive_link here, and also I upload a mask file for better masking the foreground. By the way, I share a script written by me for nerfdata with transforms.json colmap sparse, I wish this could be helpful for you. I see some of your issue people ask for this.😁 I love your work and I wish utilize your export mesh to perform in DR pipeline!