search

HKUST-Aerial-Robotics / Stereo-RCNN

Code for 'Stereo R-CNN based 3D Object Detection for Autonomous Driving' (CVPR 2019)
MIT License
690 stars 177 forks source link

Bad evaluation result on validation set [Solved] #8

Closed Jokoe66 closed 5 years ago

Jokoe66 commented 5 years ago

I got this bad evaluation result on val set, as the following:

Thank you for participating in our evaluation!
Loading detections...
number of files for evaluation: 3446
  done.
save result//plot/car_detection.txt
car_detection AP: 18.119764 17.805336 17.665022
save result//plot/car_orientation.txt
car_orientation AP: 17.903425 17.490780 17.163260
save result//plot/car_detection_ground.txt
car_detection_ground AP: 11.219302 11.579988 9.090909
save result//plot/car_detection_3d.txt
car_detection_3d AP: 9.090909 9.090909 9.090909
Your evaluation results are available at:
result/

@PeiliangLi I use cuda8.0 and pytorch0.3.0 to run the model. The result .txt files are maintained below. trainlog.txt contains training log of the first 8 epochs. Is there any idea about this situation? Many thanks. result.zip trainlog.txt

PeiliangLi commented 5 years ago

Did you write the result using our test_net.py? It seems that each of your result files only contains one object. That's the reason.

Jokoe66 commented 5 years ago

Did you write the result using our test_net.py? It seems that each of your result files only contains one object. That's the reason.

yes, I ran sh test.sh

Jokoe66 commented 5 years ago

Did you write the result using our test_net.py? It seems that each of your result files only contains one object. That's the reason.

using your trained model, I can get relatively normal result.

Jokoe66 commented 5 years ago

Did you write the result using our test_net.py? It seems that each of your result files only contains one object. That's the reason.

Oh, sorry to see that I edited kitti_utils.write_detection_results and got the write mode as w instead of a.

PeiliangLi commented 5 years ago

ok, please tell me if you get results using your trained model.

Jokoe66 commented 5 years ago

ok, please tell me if you get results using your trained model.

Yes, it's the reason. I can get results close to yours. Thanks.