search

kaist-avelab / K-Radar

4D Radar Object Detection for Autonomous Driving in Various Weather Conditions
309 stars 44 forks source link

Weird radar point cloud #1

Closed shawrby closed 9 months ago

shawrby commented 1 year ago

Thank you for your great research.

I processed the radar cube with CFAR and extracted the radar point cloud as shown in the figure below.

image

Is this radar point cloud correct?

DongHeePaek commented 1 year ago

Hello, @shawrby

We appreciate for you for your interest for our work.

The hyper-parameters for CFAR has not been optimized, since we utilize 'quantile' function for the network input.

We recommend you to adjust the hyper-parameters for CFAR by comparing the Radar point cloud with the calibrated Lidar point cloud.

The function is on line 673 of https://github.com/kaist-avelab/K-Radar/blob/f561bb5373e6d5aa85309310553ff2d63440c506/datasets/kradar_detection_v1_0.py

In addition, the hyper-parameters for CFAR are on line 130 of https://github.com/kaist-avelab/K-Radar/blob/main/configs/cfg_RTNH.yml.

Thank you.

Toytiny commented 1 year ago

Hi, @shawrby, we are also trying to generate decent point clouds from radar tensor. Have you got any ideal results after adjusting hyper-parameters for CFAR?

DongHeePaek commented 9 months ago

For the latest updates for 4D Radar point cloud, please refer to following two materials:

  1. 'Example of 4D Radar point cloud data' section & 2. vis_in_open3d function in 'kradar_detection_v2_0.py'.