Qualitative Results
Single Sequence (Sparse, Posed, 256 Gauss)
Train Views: PSNR = 16.00, SSIM = 0.72, LPIPS = 0.33
Test Views: PSNR = 15.78, SSIM = 0.73, LPIPS = 0.33
Left: Test-View Rendering | Right: Ground Truth
Single Sequence (Dense, Unposed, 256 Gauss)
Train Views: PSNR = 17.33, SSIM = 0.73, LPIPS = 0.31
Test Views: PSNR = 17.59, SSIM = 0.75, LPIPS = 0.30
Left: Test-View Rendering | Right: Ground Truth
Single Sequence (Dense, Posed, 2048 Gauss)
Train Views: PSNR = 19.31, SSIM = 0.77, LPIPS = 0.26
Test Views: PSNR = 19.0, SSIM = 0.77, LPIPS = 0.26
Left: Test-View Rendering | Right: Ground Truth
Single Sequence (Sparse, Unposed, 256 Gauss)
Train Views: PSNR = 20.03, SSIM = 0.76, LPIPS = 0.28
Test Views: PSNR = 19.8, SSIM = 0.76, LPIPS = 0.29
Left: Test-View Rendering | Right: Ground Truth
Multi Sequence (5 Sequences, Dense, Posed, 256 Gauss)
Train Views: PSNR = 15.15, SSIM = 0.71, LPIPS = 0.36
Test Views: PSNR = 15.03, SSIM = 0.70, LPIPS = 0.34
Left: Test-View Rendering | Right: Ground Truth
Multi Sequence (5 Sequences, Sparse, Unposed, 256 Gaussians)
Train Views: PSNR = 13.75, SSIM = 0.69, LPIPS = 0.35
Test Views: PSNR = 13.71, SSIM = 0.69, LPIPS = 0.35
Left: Test-View Rendering | Right: Ground Truth
From the above results, we find that our model can replicate 3DGS results in the dense and posed setting for a single sequence (object). We are also able to perform near equivalently in more difficult settings with little decrease in accuracy including the most difficult setting of sparse and unposed. We also show preliminary results of our model exhibiting the ability to generalize across sequences which Vanilla 3DGS can’t do. While the results are blurry here, it is important to note that our model is trained only on 5 different sequences here and scaling to the complete category across the CO3D dataset will help the model learn better 3D priors. Moreover, by scaling up the number of Gaussians, which we discuss in the next section, we will start seeing improved renderings for the generalizable case as well.
Quantitative Results
| Model / Metrics | PSNR | SSIM | LPIPS |
| Vanilla GS Dense | 18.61 | 0.79 | 0.28 |
| Vanilla GS Sparse | 16.69 | 0.61 | 0.40 |
| Ours Sparse | 15.78 | 0.73 | 0.33 |
| Ours Unposed | 17.59 | 0.75 | 0.30 |
| Ours Sparse + Unposed + Generalizable | 13.55 | 0.70 | 0.35 |
Ablations
Q. What improves the performance of Gaussian Splatting?
A. Number of Gaussians
Left: 256 Gaussians | Middle: 512 Gaussians | Right: 1024 Gaussians
Left: 2048 Gaussians | Middle: 4096 Gaussians | Right: 8192 Gaussians
Left: 16k Gaussians | Middle: 32k Gaussians | Right: 64k Gaussians
We did an ablation to find out what affects the results of vanilla Gaussian Splatting as we consider it as our upper bound. We found out that the number of Gaussians has the most effect on the quality of the rendering results. Above, we do a test for an object with varying numbers of Gaussians and find that 8k Gaussians is probably the best tradeoff. We also found that this is consistent across different object categories.