why is the result of kitti different from the paper ？especially for Cyclist

[jaminChen999]

Kitti Car 90.46 % 88.83 % 74.76 % ，Pedestrian 83.70 % 73.62 % 68.28 % ，Cyclist 82.34 % 74.45 % 64.91 %
Paper Car 90.03 % 89.02 % 76.11 % ，Pedestrians 83.92 % 73.70 % 68.31 % ，Cyclist 84.06 % 75.46 % 66.07 %

[jaminChen999]

@zhaoweicai

[zhaoweicai]

Do you use the pretrained models or the one trained by yourself? The pretrained models should produce very close results as in the paper. If they are trained by yourself, they look pretty close to me and the training is robust. For cyclist results, please see the disclaimer

[busyboxs]

In the kitti official web, you can find this note.

Note 1: On 25.04.2017, we have fixed a bug in the object detection evaluation script. As of now, the submitted detections are filtered based on the min. bounding box height for the respective category which we have been done before only for the ground truth detections, thus leading to false positives for the category "Easy" when bounding boxes of height 25-39 Px were submitted (and to false positives for all categories if bounding boxes smaller than 25 Px were submitted). We like to thank Amy Wu, Matt Wilder, Pekka Jänis and Philippe Vandermersch for their feedback. The last leaderboards right before the changes can be found here!