Closed gap705 closed 7 months ago
Hi @gap705, I have the following suggestions according to your questions: (1) I think you should synthesize training data with a random depth range covering your test scenario. (2) The saved inverse depth map in our task is less than 2.0 (real depth > 0.5), so I multiply 2^15 here for depth map saving. So, please check the value of your generated data to see whether the overflow problem happens (real depth < 0.5). (3) After checking the data as in (2), I think a smaller depth challenges the two-view stereo depth estimation. As an example, we find forward hand movement degrades the rendering performance when GPS-Gaussian is applied in half-body real-time communication systems.
Thanks for your reply and help.
Hello, I set the radius to 0.8m. When saving the depth map, I multiplied it by 2**12 according to your suggestion and the depth map is correct. But at this time, the results of the verification set in stage 2 will be unclear. Is it because the depth value precision is not enough? Thanks.
I think 2^14 is enough for your case. The blurred results may caused by many factors, I think you can save the point clouds un-projected by binocular depth estimation to check whether the two partial point clouds are consistent. If a significant mismatch is witnessed, the problem should come from depth estimation or the precision of GT depth data saving.
Thanks again for your help.
Thank you for your great work. I would like to ask a question. When generating a depth map based on THuman2.0, I set the radius to 0.8m. At this time, the depth value at a closer position will be wrong, but the RGB map is correct. What is the reason for this situation? How should I handle this? Thanks.