This repo contains our group project for Special Topics in AI: Deep Learning taught by Professor Qi (Rose) Yu in Spring 2019. Our project builds off the Learning to See in the Dark paper by Chen et. al. (2018). Their paper introduces an end-to-end deep learning model which converts a picture taken in the dark to one the is taken in ideal lighting conditions.

Our contribution has been to simulate dark conditions on readily available datasets, rather than using professinal photography equipment and generating relatively few images. Essentially, we are showing a way forward that achieves similar results to Chen et. al. (2018) but much more cheaply. Modifying readily available datasets, data augmentation, to achieve similar or better accuracy allows us to work around the data collection bottleneck currently experienced by Deep Learning practitioners. We often see a correlation between using more data and achieving better model accuracy. Given this correlation, we can infer that data collection is expensive, therefore, model accuracy is also expensive. We must endeavor to remove or reduce the data collection bottleneck to improve the adoption of Deep Learning technologies by industry.

Our project was executed in a proposal, milestone 1, milestone 2, and final report stage. Our folder structure relates to these stages.

1. Literature review: lit_review
2. Replication of Chen et. al. (2018): Learning-to-See-in-the-Dark
3. Identifying relevant datasets
   • ImageNet
   • raise dataset
4. Proposal report: proposal

1. Replication results and benchmarking: replication

1. Data augmentation of relevant datasets: simulating_low_light
2. Results generation against relevant datasets: milestone2

1. lastmile
   • Results generation for each sub-problem
   • Results from Zero-Shot Learning
   • Results from Transfer Learning
   • Presentation of Results

Recommended hardware for replicating Chen et. al.

• Ubuntu + Intel i7 CPU + Nvidia Titan X (Pascal) with Cuda (>=8.0) and CuDNN (>=5.0)

Recommended hardware for replicating our paper.

• 1 Tesla K80; p2.xlarge instance from AWS will be sufficient. The AWS Deep Learning AMI (Ubuntu), ami-060865e8b5914b4c4, will help with installation.

Both our paper and Chen et. al. (2018) use Tensorflow. Chen et. al. (2018) uses the slim framework, a contribution module in Tensorflow. We used Keras. Please note that Keras is being established as the high-level API for Tensorflow 2.0. Both approaches require using lower-level Tensorflow for different stages of the model.

Please use CS7180 GitHub Issues for any questions or improvements.