Multitask Learning for Object Localization With Deep Reinforcement Learning

In object localization, methods based on a top-down search strategy that focus on learning a policy have been widely researched. The performance of these methods relies heavily on the policy in question. This paper proposes a deep Q-network (DQN) that employs a multitask learning method to localize class-specific objects. This DQN agent consists of two parts, an action executor part and a terminal part. The action executor determines the action that the agent should perform and the terminal decides whether the agent has detected the target object. By taking advantage of the capability of feature learning in a multitask method, our method combines these two parts by sharing hidden layers and trains the agent using multitask learning. A detection dataset from the PASCAL visual object classes challenge 2007 was used to evaluate the proposed method, and the results show that it can achieve higher average precision with fewer search steps than similar methods.