PRBonn/lidar-bonnetal

Inference, and training not working on Semantic KITTI dataset.

AbhishekKaushikCV opened this issue · 2 comments

  • I cloned the repository, downloaded the Semantic KITTI dataset, downloaded the pretrained models.

  • The dataset structure is as follows:
    data_structure

  • I installed the mentioned packages, and libraries in requirements.txt.

  • The visualization of ground truth is working without any error. I used this command to visualize the ground truth.
    $ ./visualize.py -d /work/akmt/data/kitti -s 00

  • I am able to generate the labels on the KITTI dataset using pretrained models darknet21, and darknet53 without any error. The weights are uploading succesfully without any error. I used this command:
    ./infer.py -d /work/akmt/data/kitti -l /work/akmt/lidar-bonnetal/train/tasks/semantic/pretrained_model_predictions -m /work/akmt/lidar-bonnetal/train/tasks/semantic/pretrained_models/darknet21

  • For Model's predictions visualization, used this command:
    ./visualize.py -p /work/akmt/lidar-bonnetal/train/tasks/semantic/pretrained_model_predictions/darknet21 -d /work/akmt/data/kitti -s 00

  • But the result is very poor on different architectures as well, after some colorful predictions, the model is predicting only one label unlabelled, and therefore the output is black after some scans:
    darknet53_on_kitti

  • I am not able to find the cause of this behavior.

  • If i calculate the IOU, then all the IOUs are zero: used this command:
    ./evaluate_iou.py -p /work/akmt/lidar-bonnetal/train/tasks/semantic/pretrained_model_predictions/darknet21 -d /work/akmt/data/kitti --split valid

  • For training, there are some warnings but the the training process starts, but the update, loss, acc, iou all are zeros. I tried with and without pretrained models, but getting same zeros. using this command:
    ./train.py -d /work/akmt/data/kitti -ac /work/akmt/lidar-bonnetal/train/tasks/semantic/config/arch/darknet21.yaml -dc /work/akmt/lidar-bonnetal/train/tasks/semantic/config/labels/semantic-kitti.yaml -l /work/akmt/lidar-bonnetal/train/tasks/semantic/training_logs -p /work/akmt/lidar-bonnetal/train/tasks/semantic/pretrained_models/darknet21
    training_zeros

  • Did somebody face similar kinda problem, If so can you please help me to solve these problems. I don't understand if these problems are due to environment setup or coding.

Thanks in advance!!!

I have the same problem and I dont know why..

Did you slove it?