Project for PhD Course: A Comparative Introduction to Deep Learning Frameworks: TensorFlow, PyTorch and JAX 2021-22
In this project, we implemented a stereo matching network that exploits stack hourglass network for cost aggregation. This architecture achieves high performance on depth prediction on KITTI benchmark.
We utilize the training split of KITTI-2015 for training and evaluation.
The stereo matching network is pretrained on the frist 160 images of KITTI-2015 for 300 epochs, the pre-trained checkpoint can be found at ckpt sub-folder.
Specifically, the script for pre-training in PyTorch could be run with following command:
python torch_train.pyBesides, the TensorFlow version could be run with:
CUDA_VISIBLE_DEVICES=0 python tf_train.pyIf you have any issue or question, email to Youmin Zhang: youmin.zhang2@unibo.it