search

Owen-Liuyuxuan / visualDet3D

Official Repo for Ground-aware Monocular 3D Object Detection for Autonomous Driving / YOLOStereo3D: A Step Back to 2D for Efficient Stereo 3D Detection
https://owen-liuyuxuan.github.io/papers_reading_sharing.github.io/3dDetection/GroundAwareConvultion/
Apache License 2.0
362 stars 77 forks source link

Are the detection results is on test split or validation split? #32

Closed jichaofeng closed 2 years ago

jichaofeng commented 2 years ago

I find the detailed detection results on web page : https://github.com/Owen-Liuyuxuan/visualDet3D/releases/tag/1.1. I wander to which data split is the detection results based on. Thank you.

Owen-Liuyuxuan commented 2 years ago

That was trained on the "test_split" and evaluate on the KITTI server.

jichaofeng commented 2 years ago

Thank you for your answer,Do you have the detection results on validation split ?

Owen-Liuyuxuan commented 2 years ago

The result for the YOLOStereo3D is rather stable to reproduce.

I only store the test split result for documentation and demos. If it is needed I can retrain one.

jichaofeng commented 2 years ago

Thank you, I will reproduce the results on validation split to compare with my method.

Owen-Liuyuxuan commented 2 years ago

You can also checkout the paper.

According to my record, the full result of the published run is:

Car AP(Average Precision)@0.70, 0.70, 0.70:
bbox AP:97.38, 89.47, 69.58
bev  AP:82.48, 57.61, 44.05
3d   AP:72.06, 46.58, 35.53
aos  AP:95.59, 86.72, 67.38
Car AP(Average Precision)@0.70, 0.50, 0.50:
bbox AP:97.38, 89.47, 69.58
bev  AP:96.47, 77.39, 60.26
3d   AP:93.88, 74.39, 59.46

Many other runs of the baseline method produce better/similar results (45-47+), but you could treat this as the one you used to compare.

jichaofeng commented 2 years ago

Thank you for your help.