We evaluate the contribution of each of the proposed augmentations on the SurfaceNet model.
The model is trained on the MatSynth dataset for 400,000 steps (approx. 96 hours), with batch size of 8
and learning rate of 2e-5. We enable the adversarial discriminator after 100,000 steps.
Evaluation is conducted in terms of RMSE for all maps (except the Normal map for wich we compute the cosine
distance).
Results clearly show the improvement introduced by each of the augmentation, thus justifying their use in
the creation of the dataset.
Metric | Diffuse | Normal (Cos dist) | Roughness | Specular |
---|---|---|---|---|
No Augmentation | 0.137 | 0.672 | 0.184 | 0.088 |
Crop Only | 0.133 | 0.664 | 0.181 | 0.085 |
Crop + Rotation | 0.133 | 0.661 | 0.180 | 0.078 |
Crop + Rotation + Relightings | 0.093 | 0.536 | 0.159 | 0.054 |