About PSNR in the scene Bricks of Ricoh360

[JiaoChenGuang-JiaoHan]

I am sorry that I was unable to reproduce the PSNR of the bricks scene in the paper (24.62 in the paper) and there is a significant gap (ours 22.38). I would like to know if there are any issues during the training process. Could you help me with this?

❯ cd /home/disk1/jh/ODGS ; /usr/bin/env /bin/python3 /home/zzj/.vscode-server/extensions/ms-python.debugpy-2024.12.0/bundled/libs/debugpy/adapter/../../debugpy/launcher 45193 -- /home/disk1/jh/ODGS/train.py -s /home/disk1/jh/ODGS/dataset/Ricoh360_ODGS/bricks -m output/Ricoh360/brick/0221/test_1 --eval
Optimizing output/Ricoh360/brick/0221/test_1
Output folder: output/Ricoh360/brick/0221/test_1 [21/02 20:21:23]
Reading Transforms from OpenMVG [21/02 20:21:23]
# of Train: 50, # of Test: 50 [21/02 20:21:31]
Points without camera position (Green points) are initialized [21/02 20:21:31]
Loading Training Cameras [21/02 20:21:31]
Loading Test Cameras [21/02 20:21:32]
Number of points at initialisation : 5958 [21/02 20:21:33]
Training progress: 3%|█ | 1000/30000 [00:50<19:15, 25.11it/s, Loss=0.1385169, #pts=21457]
[ITER 1000] Eval test: L1 0.06728 PSNR 18.839 SSIM 0.5840 LPIPS 0.5278 #Points 21457 [21/02 20:22:29]
Training progress: 23%|██████▎ | 7000/30000 [12:55<1:07:29, 5.68it/s, Loss=0.0850651, #pts=1053220]
[ITER 7000] Eval test: L1 0.04119 PSNR 22.193 SSIM 0.7254 LPIPS 0.2488 #Points 1053220 [21/02 20:34:50]

[ITER 7000] Saving Gaussians [21/02 20:34:50]
Number of Nan positions before saving ply: 0 [21/02 20:34:50]
Training progress: 33%|████████▊ | 9800/30000 [23:41<1:08:42, 4.90it/s, Loss=0.0846332, #pts=1320249]
[ITER 9800] Eval test: L1 0.03979 PSNR 22.494 SSIM 0.7411 LPIPS 0.2249 #Points 1320249 [21/02 20:45:36]

[ITER 9800] Saving Gaussians [21/02 20:45:36]
Number of Nan positions before saving ply: 0 [21/02 20:45:36]
Training progress: 100%|██████████████████████████| 30000/30000 [2:41:09<00:00, 3.10it/s, Loss=0.0855779, #pts=2094617]

[ITER 30000] Eval test: L1 0.04086 PSNR 22.383 SSIM 0.7316 LPIPS 0.2218 #Points 2094617 [21/02 23:03:15]

[ITER 30000] Saving Gaussians [21/02 23:03:15]
Number of Nan positions before saving ply: 0 [21/02 23:03:15]
Learning Gaussian 9718.931344270706 [21/02 23:03:21]