Leap: a Model-Based Reinforcement Learning Framework for Fast Object Detection

Create a new algorithm for Object DetecKon through Python. The starKng point of the project was to consider the model-free algorithm described in Urbana-Champaign’s paper where they showed an algorithm for Object DetecKon based on Reinforcement Learning. The idea was to create an agent that could start from a state (a bounding box in our case) and reach the final target. The goal of our project was to make this project get closer to real-Kme simulaKons. Seen the projects for Fast RCNN and Faster RCNN, the goal of our project was to create an Object DetecKon algorithm based on Reinforcement Learning that could bypass the CNN somehow. We created a model-based algorithm, so during the Q-training, we recorded the TransiKons (state-acKon-next_state). These data are then used for the training process of a further network. The purpose of this new network was to predict the future TransiKons, in order to bypass the CNN every acKon the Agent was taking. Ader the correct design, the correct E-greedy policy assumpKon and the correct pre-processing, we end up with a faster algorithm with a low loss in accuracy (the more predicKons you do, the worse it gets, but the faster it goes).