During the semester of winter 2018, I’ve taken a deep learning course that covers fully connected and convolutional networks, state-of-the-art architectures and regulations mechanisms, recurrent networks, natural language processing, GAN, and miscellaneous topics. For the final assignment, we have chosen to construct our own dataset by scraping Unsplash website and to infer the EXIF metadata of the images (final paper and image below). The final paper is not published. We have learned so much, particularly that the learning can only be as good as the dataset.
If you have an interest in the 19000 images dataset or in the code, feel free to send me an email.
Final assignment summary:
We used deep convolutional networks to estimate the camera settings, lens aperture, ISO sensibility and exposure time, known as the EXIF metadata solely based on the pixels of the photos. The training has been performed on a novel dataset of 19 000 high quality photos labeled with the camera settings. The results indicated that deep convolutionnal networks have trouble solving the task and achieving high accuracy. We find that the general low performances we obtained are due to the dataset corruption, some class imbalances and possibly the lack of information in the picture alone.