lnayak





Box center: (0.25, 0.25, 0.00). Side lengths: (2.01, 1.50, 1.50).


View dependence lets highlights change with the camera direction, but can also overfit the training views. The table compares performance on held-out views.
| Run | Epochs | Steps | Test PSNR (dB) | Test SSIM | Step (s) | Render/view (s) |
|---|---|---|---|---|---|---|
| Position only | 250 | 25000 | 22.950 | 0.8411 | 0.0102 | 0.0811 |
| Position and direction | 250 | 25000 | 23.502 | 0.8533 | 0.0106 | 0.0863 |


Second run minus first run: test PSNR (dB): +0.552; test SSIM: +0.0123; seconds/step: +0.0004; seconds/rendered view: +0.0051.
The coarse network guides extra samples toward occupied regions. The fine network uses these samples to recover detail, at the cost of another network pass.
| Run | Epochs | Steps | Test PSNR (dB) | Test SSIM | Step (s) | Render/view (s) |
|---|---|---|---|---|---|---|
| Uniform (Q3) | 250 | 25000 | 24.242 | 0.8866 | 0.0101 | 0.0791 |
| Coarse/fine | 250 | 25000 | 24.662 | 0.8945 | 0.0188 | 0.1522 |
Uniform: 128 samples per ray. Coarse/fine: 64 coarse and 128 fine samples per ray.

Second run minus first run: test PSNR (dB): +0.420; test SSIM: +0.0078; seconds/step: +0.0087; seconds/rendered view: +0.0731.
Alpha controls density: a larger value makes the volume more opaque. Beta controls the transition around the surface. High beta produces a thicker, softer boundary; low beta produces a sharper boundary.
High beta is usually easier to train because gradients reach samples farther from the surface. Low beta can give a more accurate surface, but needs sufficiently dense samples near it.


This run uses alpha=10.000, beta=0.020 and Eikonal weight=0.050. Learning rate: 0.0005; 500 epochs, 1024 rays per step, 128 samples per ray. The smooth density transition helps optimization, while the Eikonal loss encourages unit-length SDF gradients.
| Initialization | Beta | Eikonal weight | Epochs | Val PSNR (dB) | Val SSIM |
|---|---|---|---|---|---|
| Random | 0.05 | 0.02 | 250 | 21.224 | 0.7704 |
| Geometric | 0.05 | 0.02 | 250 | 21.033 | 0.7690 |
| Geometric | 0.02 | 0.05 | 500 | 21.494 | 0.7965 |
| Random | 0.02 | 0.05 | 250 | 21.285 | 0.7796 |
| Random | 0.1 | 0.02 | 250 | 20.735 | 0.7339 |
I initialized the distance network to approximate a sphere, then trained all its weights. This avoided the distant surfaces that appeared with random initialization. I selected the result using validation scores and the rendered surface.

| Run | Iterations | Gaussians | Covariance | Split | Views | PSNR (dB) | SSIM |
|---|---|---|---|---|---|---|---|
| Truck | 1000 | 10000 | isotropic | test | 5 | 29.275 | 0.9388 |
| Parameter | Initial learning rate | Final learning rate |
|---|---|---|
| opacities | 0.05 | 0.05 |
| scales | 0.005 | 0.005 |
| colours | 0.0025 | 0.0025 |
| means | 0.00016 | 0.00016 |
Loss: L1. Learning-rate schedule: constant.






| Run | Iterations | Gaussians | Covariance | Split | Views | PSNR (dB) | SSIM |
|---|---|---|---|---|---|---|---|
| Materials baseline | 1000 | 10000 | isotropic | validation | 100 | 17.764 | 0.6505 |
| Parameter | Initial learning rate | Final learning rate |
|---|---|---|
| opacities | 0.05 | 0.05 |
| scales | 0.005 | 0.005 |
| colours | 0.0025 | 0.0025 |
| means | 0.00016 | 0.00016 |
Loss: L1. Learning-rate schedule: constant.


| Run | Iterations | Gaussians | Covariance | Split | Views | PSNR (dB) | SSIM |
|---|---|---|---|---|---|---|---|
| Materials modified | 3000 | 10000 | anisotropic | validation | 100 | 23.602 | 0.8635 |
| Parameter | Initial learning rate | Final learning rate |
|---|---|---|
| opacities | 0.05 | 0.05 |
| scales | 0.005 | 0.0005 |
| colours | 0.01 | 0.01 |
| means | 0.003 | 0.0003 |
| quats | 0.001 | 0.001 |
Loss: L1. Learning-rate schedule: exponential: means and scales decay to 0.1x by final update.


| Setting | Baseline | Modified |
|---|---|---|
| Isotropic | True | False |
| Iterations | 1000 | 3000 |
| Learning-rate schedule | constant | exponential: means and scales decay to 0.1x by final update |
| Initial position standard deviation | 0.2000 | 0.7000 |
| Initial opacity | 0.9997 | 0.1000 |
Modified minus baseline: PSNR +5.838 dB; SSIM +0.2131. The modified run has the higher PSNR. Anisotropic Gaussians can stretch along different directions. Learning-rate decay makes the final position and scale updates smaller. Larger position updates help points reach the objects. A higher colour learning rate helps their appearance converge. The modified run also trains for more iterations. I spread the initial random points more widely to cover the scene. Lower initial opacity lets points behind the foreground points contribute during early training.
The modified run separates the objects much more clearly than the baseline. Some reflections remain softer than the ground truth; each Gaussian still has one RGB colour for all viewing directions.