search

tsinghua-rll / VoxelNet-tensorflow

A 3D object detection system for autonomous driving.
MIT License
451 stars 123 forks source link

what is the performance of trained model #17

Open qianguih opened 6 years ago

qianguih commented 6 years ago

Hi,

Thanks for the amazing work on the implementation. I am wondering if it is possible to share some performance numbers of the trained model? Thanks!

HectorAnadon commented 6 years ago

Same question here because I am a bit unsure I did something wrong. I trained it with the default parameters and all training data (evaluation method https://github.com/smallcorgi/3D-Deepbox/tree/master/kitti_eval):

number of files for evaluation: 7481
  done.
save /home/hleon/Documents/VoxelNet-tensorflow/data/results/data/plot/car_detection.txt
car_detection AP: 26.815498 29.592041 28.789532
save /home/hleon/Documents/VoxelNet-tensorflow/data/results/data/plot/car_orientation.txt
car_orientation AP: 11.783903 12.759486 12.466972
save /home/hleon/Documents/VoxelNet-tensorflow/data/results/data/plot/car_detection_ground.txt
car_detection_ground AP: 63.132660 70.432945 69.036919
Eval 3D bounding boxes
save /home/hleon/Documents/VoxelNet-tensorflow/data/results/data/plot/car_detection_3d.txt
car_detection_3d AP: 45.327732 46.881519 47.241013

This is tested against all training data of 3D kitti dataset.

I have noticed that this is only implemented for cars, or at least I only get cars as predictions.