KITTI training performance
zxc-2015 opened this issue · comments
Hi packyan,
The KITTI val performance of using the pre-trained model is very nice, which is:
Average Precisions:
- Class '0' - AP: 0.8874476379466757
- Class '1' - AP: 0.8776504929104587
- Class '2' - AP: 0.9094174650326788
- Class '3' - AP: 0.7935572902584845
- Class '4' - AP: 0.6938401233821029
- Class '5' - AP: 0.7484522134591276
- Class '6' - AP: 0.9765143493123802
- Class '7' - AP: 0.6567834290009646
mAP: 0.8179578751628592
However, when doing the training with the Imagenet pretrained-model, I am only able to get the following results, which is:
Average Precisions:
- Class '0' - AP: 0.5895481840925929
- Class '1' - AP: 0.056750284687503164
- Class '2' - AP: 0.05041402945683988
- Class '3' - AP: 0.25446338964588194
- Class '4' - AP: 0.0
- Class '5' - AP: 0.042098299062717656
- Class '6' - AP: 0.0
- Class '7' - AP: 0.0
mAP: 0.12415927336819195
�
I'm wondering what thing do I miss or configure? I use the same label-transforming pre-processing code as in the instruction and training without changing the python arguments. Is anything wrong I did or any arguments need to be changed?
Thanks a lot!
Imagenet pretrained-model is just for darknet layer, the latter layers you should train by kitti dataset.
Average Precisions:
Car + Class '0' - AP: 0.8874329809457615
Van + Class '1' - AP: 0.8776504929104587
Truck + Class '2' - AP: 0.9094174650326788
Pedestrian + Class '3' - AP: 0.7935575396672143
Person_sitting + Class '4' - AP: 0.6938401233821029
Cyclist + Class '5' - AP: 0.7484521975616071
Tram + Class '6' - AP: 0.9765143493123802
Misc + Class '7' - AP: 0.656781273677661
mAP: 0.817955802811233
Average Precisions:
Car + Class '0' - AP: 0.5925623770514821
Van + Class '1' - AP: 0.06155695621575082
Truck + Class '2' - AP: 0.046829029317335555
Pedestrian + Class '3' - AP: 0.21473984288447187
Person_sitting + Class '4' - AP: 0.0
Cyclist + Class '5' - AP: 0.04053911530526016
Tram + Class '6' - AP: 0.0
Misc + Class '7' - AP: 0.0
mAP: 0.11952841509678755
Hi zxc-2015, I encountered the same problem as you.
I divided the kitti 7481 training images into 3712(train)+3769(val)
Did you solve this problem?
I am trying the whole training set to train the model, and will see the result after 100 epoch
Me too, did you solve it?