search

NVIDIA-AI-IOT / CUDA-PointPillars

A project demonstrating how to use CUDA-PointPillars to deal with cloud points data from lidar.
Apache License 2.0
502 stars 148 forks source link

The performance specified in github is not reproduced. #80

Open KwangjinChoi opened 1 year ago

KwangjinChoi commented 1 year ago

Hi, I would like to reproduce the preformance of github with the pre-trained model given by CUDA-PointPillars. However, I would like to ask help because it is not reproduced.

Compared to the pytorch model, the preformance of the onnx converted model is reduced by 10%. Is there someone who is in a similar situation to me or can you help me?

Below is the preformance that I reproduced.

                  |car    | Pedestrain    |Cyclist
CUDA-PointPillars | 67.67 | 50.40         | 56.73
OpenPCDet         | 77.28 | 52.29         | 62.68
wayyeah commented 1 year ago

I have the same problem, have you solved it? Below is the performance that I reproduced.

                                 |car    | Pedestrain    |Cyclist
CUDA-PointPillars(fp16)          | 67.50 | 50.93         | 55.86
CUDA-PointPillars(fp32)          | 67.49 | 51.17         | 56.04
OpenPCDet                        | 77.09 | 51.39         | 62.46
ACFIRSTONE commented 7 months ago

Me too,i I have the same problem!

byte-deve commented 7 months ago

Hi, are you using original point-cloud bin as CUDA-PointPillars input? Need to prune the bin and only keep points in camera FOV to reproduce the numbers.

ACFIRSTONE commented 7 months ago

Hi, are you using original point-cloud bin as CUDA-PointPillars input? Need to prune the bin and only keep points in camera FOV to reproduce the numbers.

How to prune the bin?

I know, I wrote a script to remove the points where x<0