The KITTI Vision Benchmark Suite

We enhance the well-known Faster R-CNN method by adding viewpoint inference capabilities. Convolutional features are computed and shared for use in the tasks of region proposal, classification, and orientation estimation, for efficiency reasons.

Scale=500, 300 RPN proposals, 3 anchor scales and 3 anchor ratios, \N_b=8, \l_r=0.001, 150k iterations.

@inproceedings{GuindelITSM,
author = {Guindel, Carlos and Martin, David and Armingol, Jose M.},
booktitle = {IEEE Intelligent Transportation Systems Magazine},
title = {{Fast Joint Object Detection and Viewpoint Estimation for Traffic Scene Understanding}},
year = {2018}
}
@inproceedings{GuindelICVES,
author = {Guindel, Carlos and Martin, David and Armingol, Jose M.},
booktitle = {IEEE International Conference on Vehicular Electronics and Safety (ICVES)},
title = {{Joint Object Detection and Viewpoint Estimation using CNN features}},
year = {2017}
}