prediction on the custom data

I have successfully run the test.py on the KITTI dataset. I am a student new to the 3D detection research. I want to ask for help, with this wonderful toolkit OpenPCDet, how can I predict using a pre-trained model on my own data provided by the teacher. Can anybody give me some help or blog which I can refer to. I want to check the result on the data my radar collected. Thanks for the kind help!

here is the bag information on my own data:

path:        2018-01-06-15-14-01.bag
version:     2.0
duration:    14:51s (891s)
start:       Jan 06 2018 15:14:01.43 (1515222841.43)
end:         Jan 06 2018 15:28:52.53 (1515223732.53)
size:        8.5 GB
messages:    45378
compression: none [8842/8842 chunks]
types:       nav_msgs/Odometry       [cd5e73d190d741a2f92e81eda573aca7]
             sensor_msgs/Imu         [6a62c6daae103f4ff57a132d6f95cec2]
             sensor_msgs/NavSatFix   [2d3a8cd499b9b4a0249fb98fd05cfa48]
             sensor_msgs/PointCloud2 [1158d486dd51d683ce2f1be655c3c181]
topics:      /fix              892 msgs    : sensor_msgs/NavSatFix  
             /imu_raw        17823 msgs    : sensor_msgs/Imu        
             /odom_encoder   17822 msgs    : nav_msgs/Odometry      
             /points_raw      8841 msgs    : sensor_msgs/PointCloud2

[alixiansen91]

Hello, on the basis of using only 3D point cloud data, how to use the reference coordinate system under KITTI? There is another question, how do you know what your coordinate system is like?

[sshaoshuai]

1. Load your custom data, and get the raw point cloud (N, 3) with x, y, z. (The provided pretrained-model used (N, 4) with x, y, z, intensity, if you doesn’t have the intensity, then you need to train a new model without intensity. )
2. You need to feed the points (N, 3 or 4) to the model, just refer to the _getitem_ function of kitti_dataset.py to repace the kitti points with your own points.

Also remember that the coordinate of your points (N, 3) should in the unified normative coordinate of PCDet as shown in the README.md. And the radar points are much sparser than the KITTI LiDAR points, so I guess you may get a bad prediction results if you only use the radar points.