Recently, I've been working with my own dataset. I realized many might be unfamiliar with how to utilize the advanced LiDAR open source tool, Openpcdet, to train their custom datasets. Hence, I've documented the steps I followed. Regardless of the type of LiDAR you use, ensure you have data in the format (x,y,z,intensity). The subsequent conversions are straightforward, ranging from formats like .txt, .pcd, .bin, to .npy.
- ✅ Build a dataloader and import your own data
- 📝 Training my own dataset (TODO)
- 📝 Test my own dataset (TODO)
ST3D/Openpcdet
├── data
│ ├── mydata
│ │ │── lidar
│ │ │── label
│ │ │── lidar
- lidar information
│ │ │── label
- label information
- For lidar annotation tools please refer this : https://github.com/Leozyc-waseda/awesome-lidar-labeling
# My dataset coordinates are:
# - x pointing to the right
# - y pointing to the front
# - z pointing up
# Openpcdet Normative coordinates are:
# - x pointing foreward
# - y pointings to the left
# - z pointing to the top
# So a transformation is required to the match the normative coordinates
points = points[:, [1, 0, 2]] # switch x and y
points[:, 1] = - points[:, 1] # revert y axis
Feel free to replace mydata with your own dataset name.
-
Configuration about your own dataset https://github.com/Leozyc-waseda/Lidar_Openpcdet_ST3D/blob/365ac6b57906a01aeac27bfbdd8f1df29b5c2744/ST3D/pcdet/datasets/mydata/mydata_dataset.py
-
Get your Lidar path and annotation path
-
Get all annotation labels informations,Regarding the annotation information, you need to change it to your own annotation information. There is only car in my json file. Since I haven't used it for training, I haven't tested this part yet.
Lidar_Openpcdet_ST3D/ST3D/pcdet/datasets/mydata/mydata_dataset.py
Lines 140 to 146 in 365ac6b
- Basically copy from kitti_dataset.yaml
- change here
python -m pcdet.datasets.mydata.mydata_dataset create_mydata_infos tools/cfgs/dataset_configs/mydata_dataset.yaml
- Training my own dataset (📝 TODO)
- Test my own dataset (📝 TODO)
This project is licensed under the MIT License.
For more details, see LICENSE file.