Over the last year I have been working on a research paper with some of my tutors at Wintec. For this paper we did research into LCFCN. You can read the paper here.

I started off with an implementation called Deepfish. I added support for Microsoft Windows. Documented and built a consistent installation script to enable reproducible installs. I also implemented a custom loss function for the network. I developed wrappers so that my custom labeling tool could generate predictions and make training data correctly.

I developed a custom point level annotation tool in Python. This enables the user to quickly label all the objects in an image and save it to an extensible text-based format. The tool used OpenCV to display the images. It was put to the test labeling more than 1000 pictures. Due to the annotation format being lightweight, sharing annotations between collaborators was easy.

Over the research we used a variety of open-source resources and on-site photography to make custom data sets. To extend the data sets data augmentation techniques were used. Due to the architecture of the network, it cannot distinguish between an image and a mirrored copy. This enabled us to 8X the data we labeled.

One of the data sets that we used we have published. It is a data set of cows labeled from aerial photography provided by the New Zealand government. All the pictures are 500 pixels square and have a density of 10 centimeters per pixel. The data set is hosted here.

I volunteered my personal computer to provide the compute power necessary. It took multiple days per dataset to train the network. Once trained the network would then be tested against a final holdout validation set. I wrote some scripts to take the final numbers and compute overall loss values. Participation certificate for ISCV2022

The paper was then presented at an international conference. I received a certificate of participation for my contributions to the paper.