chcorbi
commented
1 day ago Hi @booker-max,
We project a single frame of point clouds that aligns most closely with the corresponding image frame. However, for the "augmented" training set, which incorporates depth completion to increase label density, we aggregate data from the four preceding and four succeeding LiDAR frames before projection.
If needed, Appendix B of our arXiv preprint provides more detailed information about the projection process and data collection. And details about the depth completion pipeline are also available in Appendix C.
Hello, in your description, you mentioned that ‘the dataset was collected using two 360° cameras in a top-bottom setup and a LiDAR sensor, and it includes accurate depth and disparity labels by projecting 3D point clouds onto equirectangular images.’ I would like to ask whether you project a single frame of point clouds onto the image to obtain the depth map, or if you combine multiple frames of point clouds before projecting them to generate the depth map?