Main Approach:- With the independent feature being the image itself and the dependent feature being two set of coordinates, our approach revolves around building a standard Convolutional Neural Network. The architecture of the CNN used is ResNet50,Resnet18 (https://arxiv.org/abs/1512.03385). A custom head is added to the CNN to get the desired 4 numbers as output.
Some key techniques used to improve the model:- 1.) Data Augmentation.(Changing the brightness, contrast, rotation, zooming, warping of the given images to artificialy generate more data.) 2.) Differential Learning rates.(https://towardsdatascience.com/transfer-learning-using-differential-learning-rates-638455797f00) 3.) One cycle fitting policy (https://arxiv.org/abs/1803.09820) 4.) Cyclical Learning rates (https://arxiv.org/abs/1506.01186) 5.) Stochastic Gradient Descent with Restarts.(The basic idea is to reset our learning rate after a certain number of iterations so that we can pop out of the local minima if we appear to be stuck.)
Main libraries used:- Fastai (https://github.com/fastai/fastai) built on top of PyTorch. Numpy Pandas OpenCV