Awesome
kitti_eval
evaluate_object_3d_offline.cpp
evaluates your KITTI detection locally on your own computer using your validation data selected from KITTI training dataset, with the following metrics:
- overlap on image (AP)
- oriented overlap on image (AOS)
- overlap on ground-plane (AP)
- overlap in 3D (AP)
Compile evaluate_object_3d_offline.cpp
with dependency of Boost and Linux dirent.h
(You should already have it under most Linux).
Run the evalutaion by:
./evaluate_object_3d_offline groundtruth_dir result_dir
Note that you don't have to detect over all KITTI training data. The evaluator only evaluates samples whose result files exist.
Updates
- June, 2017:
- Fixed the bug of detection box filtering based on min height according to KITTI's note on 25.04.2017.