January - March 2019

Patch Descriptor

As part of a university module, I used de-noising and representation learning for generating a patch descriptor that is able to perform tasks such as matching, retrieval and verification. The HPatches dataset was used for benchmarking.

Among the methods explored were a shallow U-Net and a feed-forward DnCNN for the de-noiser, as well as an L2-Net for the descriptor.

I documented my findings in a final report. The code for all the models is stored in a GitHub repository.