escaped / GCNetwork

Geometry and Context Network

This project implements the GCNetwork developed by Kendal, et al(2017). Currently, we only train with the Middlebury 2014 dataset for indoor object depth localization.

Directories:

code : contains main function (main.py), core function(end_endlearning.py) and helper functions(conv3dTranspose and pfm loader)

data : stores middlebury data

log : stores log file, which is useful for visualization

model : trained model.

Running code by calling main.py. Arguments for main.py:

mode : 0 for prediction, 1 for training with existing model, 2 for training with new model

data : path for training data

-mpath : pretrained model path. Provided when mode is 0 for 1.

-bs : batch_size. default = 1

-lr : learning_rate. default = 0.001

-ep : epochs. default = 10

-mspath : model_save_path. used when mode is 1 or 2

-lspath : log_save_path. used when mode is 1 or 2. This is the log file used in Tensorboard.

-vdata : path for validation path

-pspath : file for saving predicted result. Used when mode is 0.

ex: srun --pty python code/main.py 2 data/mb_data/mb_train.npz \-mspath model/mb_model/mbModel.h5 -lspath log/mb_log/log -vdata data/mb_data/mb_val.npz --epochs

Reference: Kendall, Alex, et al. "End-to-End Learning of Geometry and Context for Deep Stereo Regression." arXiv preprint arXiv:1703.04309 (2017).