When I send the point cloud results to 3D Gaussian Splatting for training , the effect is not very good. Can you help me analyze it?

[crmlei]

(1)I have runned the program on the tandt/truck dataset( 251 images ) using default parameters:

python ace_zero.py "/dataset/tandt_db/tandt/truck/images/*.jpg" ./output --export_point_cloud True

(2)and export point cloud data using export_point_cloud.py:

python export_point_cloud.py ./output/point_cloud_out.txt --network ./output/iteration25.pt --pose_file ./output/poses_final.txt

(3) and translate the result to colmap format ( using codes copied from issue #30 ), get four files: points3D.ply,images.txt,points3D.txt,cameras.txt

(4) and use 3D Gaussian Splatting(https://github.com/graphdeco-inria/gaussian-splatting) to train model ：
python train.py -s /dataset/colmap/truck

(5) and I using 3DGaussianSplatting web viewer( viser from https://github.com/yzslab/gaussian-splatting-lightning) ,and the effect is not very good (As a comparison,the view is good when using COLMAP to generate the point cloud data )

Do I need to adjust some parameters? or this project or does ACEZERO require more images for processing? ( the truck dataset only has 251 images)

Also, is it possible that it's a problem with the viewer? (I haven't tried any other viewers yet, but when I used colmap to generate point cloud data, the final rendering result I saw was good.)

Thanks!

[crmlei]

I use another web viewer to get good result.