• Unofficial Pytorch implementation of - Unsupervised Cross-spectral Stereo Matching by Learning to Synthesize
• Steps to train on Pittsburgh dataset:
• Qualitative Results
• Quantitative Results
• Trained Weights:
• Training
• Summary
• Differences from the original Paper

Unofficial Pytorch implementation of - Unsupervised Cross-spectral Stereo Matching by Learning to Synthesize

Steps to train on Pittsburgh dataset:

• Set up the config file in ./config folder, adjust the hyperparameters as per your need
• Download the pittsburgh RGB-NIR stereo dataset specify the location in the config file (basepath)
• Don't forget to execute pip install -r requirements.txt

Qualitative Results

Left (RGB)	Fake Left (NIR)	Right (NIR)	Fake Right (RGB)	Disparity

Quantitative Results

Method	Common	Light	Glass	Glossy	Veg	Skin	Clothing	Bag	Mean
From paper (STN + SMN)	1.13	1.55	1.05	1.52	0.89	1.23	1.14	0.98	1.18
This Implementation	0.64	1.51	1.12	1.93	0.70	1.12	1.14	1.12	1.17

Comparing STN + SMN since the spectral transalation in this implementation is not F-cyclegan but the original cyclegan.

Trained Weights:

• Weights are available at following --> link.
• Download them and place in the folder where you will be saving your weights (./weights according to the default config file)
• The weights will be saved in the following a particular format, for eg. [epoch]_net_G_A.pth and latest_net_G_A.pth which signifies the latest checkpoint, you can specify the epoch you want to load weights from in the config file.

Training

• The model follows iterative optimization technique as described in the paper.
• Only CycleGAN is trained for 10 epochs, thus set warmup: True in config file
• Change warmup: False for training combined step 1,2,3,4
• For quantitative results, prepare a config file (eg. pittsburgh_test.yaml) and run python test.py --config ./configs/pittsburgh_test.yaml

Summary

• Summaries are created inside ./summary folder
• to view loss logs run tensorboard --logdir ./summary

Differences from the original Paper

• The spectral translation network is cyclegan and not F-cyclegan
• The weights of the losses are different, for some reason the network was not converging with the default loss weights for my training setup.
• The model is not optimized for best performance since I don't have the hardware that was used by the authors and rely on Colab for open source projects :)