This is the implementation of WESNet using PyTorch. Here is our paper.
For training:
- CUDA
- PyTorch
- Other requirements (such as visiualization of loss curve)
pip install -r requirements.txt
For data preprocessing:
- OpenCV
- Eigen
Download our surround-view dataset containing original fisheye images from here (password:cblf), and extract. The labels of the calibration site images consisting of the 2D pixel coordinates and corresponding 3D world coordinates of the selected corners are also provided in 'C_f.zip'.
- Labels: Before training, make sure that each lable file contains the same number of point pairs, we suggest keeping at least 15 pairs.
- Augmentation: Add the disturbance within a certain range to the original dataset for data augmentation. To prove efficiency, we implemented it in C++ (see prepare_training.cpp), and don't forget to change the directory to your own directory when reading and saving images.
python train.py --dataset_directory $TRAIN_DIRECTORY --batch_size $BATCH_SIZE --enable_visdom
TRAIN_DIRECTORY is the directory of training data generated in data preparation.
--enable_visdom is to help visiualize the loss curve in the training process.
View config.py for more argument details (num_epochs, learning rate, etc).
python inference.py --image_path $IMAGE_PATH$ --detector_weights $DETECTOR_WEIGHTS
IMAGE_PATH is the directory of testing data generated in data preparation, in which the organization of folders and names of testing images is the same as the training set.
DETECTOR_WEIGHTS is the model of trained weights.
See our project website to see some demo videos.