benhenryL
commented
1 year ago We appreciate your interest to our work! While our method (DWT + mask) outperforms spatial encoding model (only mask applied) in terms of test PSNR, but it does not lead to the faster convergence as the following figure shows. They are training curves for models which are trained for NeRF-Synthetic chair dataset. Spatial encoding approach shows test PSNR 33.81 with sparsity of 0.9614 and the proposed method shows 34.14 and 0.9722 for each. We may assume that DWT can contribute to improve the generalization of the model. Also, our method achieved higher test PSNR even with higher sparsity which aligns with our claim that DWT fits better to the proposed trainable mask, but it does not directly improve the convergence speed.
Very impressive work! I was wondering if you had the test training curves available from your training runs? The "Neural Fourier Filter Banks" paper observed that their wavelet inspired approach led to faster convergence than other spatial encoding approaches, and I would be curious whether or not this is something which hold for this approach also