Open viplix3 opened 3 years ago
hello @viplix3, thank you! Have you evaluate with a kitti tool ? Because I evaluate with https://github.com/prclibo/kitti_eval but I have a wrong result.
@SofianeB-03 yes, I was able to evaluate using the KITTI evaluation toolkit. I used this GitHub repo. I used the pre-trained weights of the complex-yolo model shared by the author of this repository as well as one I trained myself.
It would be helpful if you could elaborate in what sense are you getting the wrong results.
Thank you for your help. My results are: AP ~ 0.02 for pedestrian and cyclist and ~0.20 for car. It seems wrong (with the evaluation from evaluate.py on this repo, I have mAP ~ 0.88).
@viplix3 I used the same kitti tool to evaluate this model ComplexYolov4 with pre-trained weights on this repo (on Validation dataset kitti) but I have wrong results again : AP ~ 0.0 Ped,Cycl and AP ~ 0.10 for car. Can you share your results of the pre-trained model from this repo please?
@SofianeB-03, for the model I trained, I am getting these numbers on the KITTI test sub-set I've created from KITTI training data. This subset has not been used in the training/validation set of the model.
BEV (Easy/Moderate/Hard) Pedestrian: (71.67, 74.56, 74.74) Cyclist: (59.54, 64.67, 65.10) Car: (98.89, 96.90, 96.86)
3D (Easy/Moderate/Hard) Pedestrian: (56.27, 56.37, 56.85) Cyclist: (1.26, 0.90, 1.96)) Car: (46.17, 40.91, 43.01)
Thank you very much !
A brief overview of changes done and the files affected mentioned below.
Retaining box prediction confidence when converting the box from LiDAR coordinate frame to camera coordinate frame. File changed:
Dumping the output labels in the KITTI evaluation toolkit expected format. Files changed: