Open CBY-9527 opened 1 year ago
For simplicity, the first 64 beams of a 200-beam lidar have the same mask probability as a 64-beam lidar. The downsampling algorithm in this repo is the same as in the paper, except in this repo the mask probability is given directly in config.
Hello, I used the randomly down-sampling enhancement strategy (waymo->KITTI), but the performance is particularly poor, more than ten points lower than the performance without using randomly down-sampling. I have a question. Waymo is obtained by a 64-beam +4 200-beam LiDARs. The effective distance of the 200-beam LiDAR is 20m. The code you provided is to set a threshold to generate a beam mask, which seems unreasonable. Can you provide the downsampling code for down-sampling in the paper?
I have encountered some issues. May I know which torch version and spconv version you are using?
torch==1.10.1 spconv==2.1.21
Hello, I used the randomly down-sampling enhancement strategy (waymo->KITTI), but the performance is particularly poor, more than ten points lower than the performance without using randomly down-sampling. I have a question. Waymo is obtained by a 64-beam +4 200-beam LiDARs. The effective distance of the 200-beam LiDAR is 20m. The code you provided is to set a threshold to generate a beam mask, which seems unreasonable. Can you provide the downsampling code for down-sampling in the paper?