Open booker-max opened 1 week ago
Hi, the depth is the ray-Gaussian intersection. It is the median depth when the accumulated blending weights for a pixel excel 0.5.
You can simply use a L1 loss if you can compute the depth of a pixel from your LiDAR data.
"Hello, I have real LiDAR depth data that can be used for network supervision, but I found some issues when directly using LiDAR to supervise depth = rendering[6, :, :]. It seems that render_depth is not the true depth. I have two questions: