-
Notifications
You must be signed in to change notification settings - Fork 38
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Model is not training well #8
Comments
Please try this range: And you should turn off the use of ground planes, we do not have those (yet). |
@andraspalffy @shawrby When we use the KITTI format, it divides the dataset into Easy, Moderate and Hard. My question is that how can we get a similar AP for each class and mAP as in your paper. |
The evaluation code will be shared soon so you can do tests on your own. But basically, we did not use difficulties, what you see in the paper is the "Hard" category, using every object. I think your training issues originate from the config differences, see above. |
@andraspalffy Thank you. Referring to your advice, I will retrain the model. |
@andraspalffy How many epoches do you use in your training? |
@andraspalffy I still have the problem of model not learning well. Are there any other considerations? |
@shawrby How many epochs do you use in your training? |
@yaoshanliang I used 80 epochs of the default config file. |
Please share both your pointpillars config and dataset config and we will check. How do you defined your splits? |
@andraspalffy Model config: pointpillar.yaml
Dataset config: kitti_dataset.yaml
Is there something wrong? And dataset preparation was performed according to the OpenPCDet procedure |
This seems okay, except for the 'jslee_cfgs' config folder for base config, which I of course do not know. Are you sure that the location you refer to '../data/kitti' has the correct data, and when you generated the splits, everything went well? We will try to share a LiDAR training tutorial later this week. |
Thank you very much for the prompt reply.! |
@andraspalffy Can you provide your config for pp-radar? I used the default config in MMDetection3D: https://github.com/open-mmlab/mmdetection3d/blob/master/configs/base/datasets/kitti-3d-3class.py |
@yaoshanliang |
@shawrby Yes, the dataset root path is my radar data. |
Hi, we have some deadlines today and tomorrow, but we are working in the meantime to give you a nice manual of how to train similar models as we did (first LiDAR, then radar). We did not use the mmdetection3d library, but I can see some differences already in the config:
Till then, here is our most recent pp-radar config for OpenPCDet, @shawrby and @yaoshanliang may be interested as well.
|
@andraspalffy Firstly, thanks to provide such a meaningful dataset. Based on your provided config file, I try to reproduction pp-lidar result. I found the training result is not good. The training result and the config file are followed. The dataset config file:
The training config file:
Is there any problem with my config file leading bad training results? |
@ZixianReid |
Hi all, Thank you for your interest. I would like to emphasise that this is a dataset repo first, not a paper reproduction repo. |
@andraspalffy Looking forward to it ! |
Could you provide the Radar7PillarVFE function? Thank you. |
Hi, the function will be shared in the tutorial, I do not want to further convolve this thread with half-ready codes. |
Dear all, good news, we think we have found the cause of the poor results. TL;DR: please comment out this line (or equivalent step): Longer explanation:
Solution
We applied this fix, and trained a PP-LiDAR on "train", tested on "val", and got proper results. @yaoshanliang This will also partly fix the radar-based training. That being said, we will provide the VFE as promised. I will open another issue for it. |
@andraspalffy Thanks for checking, the performance of pp-lidar is close to the paper's result now. Shown below:
Looking forward to the pp-radar reproduced. |
Dear all, @BalazsSzekeres just pushed the evaluation tutorial! |
Dear @andraspalffy, has the VFE been public? |
@yaoshanliang Please refer to the following document. Please consider it as a starting point and not an exhaustive list of instructions. It should however cover the major modifications we made to adapt OpenPCDet to train PointPillars on our radar point cloud. |
have you reproduced the radar results using the mmdetection3d? I was not able to obtain similar results, especially in cyclist and pedestrian. |
Dear @chyohoo and all users, we kindly ask you to check issues and comments before you post a new one to help our work and to make it easier for others to find answers. Many of your questions were answered before. Let's go over them:
Hope this helps and gets you going. |
I try to do the experiments in your paper using OpenPCDet.
For the PP-Lidar experiment, the model was trained with the default config(https://github.com/open-mmlab/OpenPCDet/blob/master/tools/cfgs/kitti_models/pointpillar.yaml), but the following results were obtained.
Compared to Car, Pedestrian and Cyclist have poor results. Especially Cyclist doesn't seem to be learning at all.
What am I missing?
The text was updated successfully, but these errors were encountered: