Qonvolution: Towards Learning of High-Frequency Signals with Queried Convolutions

Novel view reconstructions for conventional 3D Gaussian Splatting (3DGS) (left) and our proposed method with adding QNN (right) with SfM initializations on the Mip-NeRF 360 dataset. Our method of adding QNN, faithfully reconstructs the scene providing higher quality renderings. Slide the handles to compare the baseline 3DGS vs. proposed 3DGS+QNN rendering.