Fig. 4

Depth Value Shifting in Multi-layer 3D Gaussian Ellipsoid Rendering. A layered approach for training the neural network is applied, resulting in a 3DGS reconstruction of the image. In this scenario, the network predicts 3D Gaussian properties for pixel c, with depth values D1, D2, and D3. These Gaussian ellipsoids’ parameters ensure correct color rendering when projected onto pixel c. However, when rendering the depth values of the Gaussian ellipsoids G1, G2, and G3, the resultant depth value Da for pixel c differs from the depth value predicted by the depth estimator. This discrepancy, when extended to each pixel, results in a completely disordered geometric rendering for the entire 3DGS model.