YOLO3D is inspired by Mousavian et al. in their paper 3D Bounding Box Estimation Using Deep Learning and Geometry. YOLO3D uses a different approach, we use 2d gt label result as the input of first stage detector, then use the 2d result as input to regressor model.

git clone git@github.com:ApolloAuto/apollo-model-yolo3d.git

cd apollo-model-yolo3d

conda create -n apollo_yolo3d python=3.8 numpy
conda activate apollo_yolo3d
pip install -r requirements.txt

here we use KITTI data to train. You can download KITTI dataset from official website. After that, extract dataset to data/KITTI.

ln -s /your/KITTI/path data/KITTI

├── data
│   └── KITTI
│       ├── calib
│       ├── images_2
│       └── labels_2

modify datasplit file to split train and val data customerly.

cd data 
python datasplit.py

modify train.yaml to train your model.

python src/train.py experiment=sample

log path: /logs
model path: /weights

modify convert.yaml file to trans .ckpt to .pt model

python convert.py

In order to show the real model infer ability, we crop image according to gt 2d box as yolo3d input, you can use following command to plot 3d results.

modify inference.yaml file to change .pt model path. export_onnx=True can export onnx model.

python inference.py \
          source_dir=./data/KITTI \
          detector.classes=6 \
          regressor_weights=./weights/pytorch-kitti.pt \
          export_onnx=False \
          func=image

• source_dir: path os datasets, include /image_2 and /label_2 folder
• detector.classes: kitti class
• regressor_weights: your model
• export_onnx: export onnx model for apollo

result path: /outputs

generate label for 3d result:

python inference.py \
          source_dir=./data/KITTI \
          detector.classes=6 \
          regressor_weights=./weights/pytorch-kitti.pt \
          export_onnx=False \
          func=label

result path: /data/KITTI/result

├── data
│   └── KITTI
│       ├── calib
│       ├── images_2
│       ├── labels_2
│       └── result

modify label_path、result_path and label_split_file in kitti_object_eval folder script run.sh, with the help of it we can calculate mAP:

cd kitti_object_eval
sh run.sh

• yolo3d-lighting
• skhadem/3D-BoundingBox
• Mousavian et al.

@misc{mousavian20173d,
      title={3D Bounding Box Estimation Using Deep Learning and Geometry}, 
      author={Arsalan Mousavian and Dragomir Anguelov and John Flynn and Jana Kosecka},
      year={2017},
      eprint={1612.00496},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}