Chaphlagical / Deblur-GS

[I3D 2024] Deblur-GS: 3D Gaussian Splatting from Camera Motion Blurred Images
https://chaphlagical.icu/Deblur-GS/
Other
354 stars 36 forks source link

Not being able to reproduce the paper quality #5

Closed pdxmusic closed 6 months ago

pdxmusic commented 6 months ago

tonight I ran a training session using this command:

python train.py -s /deblur-datasets/bench/ --eval --deblur

This is the final output: [ITER 89000] Evaluating train: L1 0.07883617728948594 PSNR 18.672360610961913 SSIM 0.5059403717517853 LPIPS 0.5323960781097412 [17/05 02:12:15]

This is how long it took Training progress: 100%|█| 90000/90000 [6:58:09<00:00, 3.59it/s, Loss=0.0423791

This is the ply result

result

Is there something I'm missing? Is it possible to know which parameters were used in the training results of the paper?

Thanks in advance!

Chaphlagical commented 6 months ago

I might change the order of the quaternion vector of the original 3dgs framework. Therefore, SIBR viewer might not works well. How about the result from render.py or render_video.py? Maybe I will change the order back when I have some free time. Sorry for the inconvenience.

pdxmusic commented 6 months ago

The video render is perfect!