Awesome

GraftNet: Towards Domain Generalized Stereo Matching with a Broad-Spectrum and Task-Oriented Feature

Dependencies:

• Python 3.6
• PyTorch 1.7.0
• torchvision 0.3.0
• VGG trained on ImageNet

Datasets:

• SceneFlow
• KITTI stereo 2015
• KITTI stereo 2012
• Middlebury v3
• ETH3D

Training Steps:

1. Train A Basic Stereo Matching Network:

python train_baseline.py --data_path (your SceneFlow data folder)

2. Graft VGG's Feature and Train the Feature Adaptor:

python train_adaptor.py --data_path (your SceneFlow data folder)

3. Retrain the Cost Aggregation Module:

python retrain_CostAggregation.py --data_path (your SceneFlow data folder)

Evaluation:

Evaluate on KITTI:

python test_kitti.py --data_path (your KITTI training data folder) --load_path (the path of the final model)

Evaluate on Middlebury-H:

python test_middlebury.py --data_path (your Middlebury training data folder) --load_path (the path of the final model)

Evaluate on ETH3D:

python test_middlebury.py --data_path (your Middlebury training data folder) --load_path (the path of the final model)

Pretrained Models:

Google Drive