Deep Active Robotic Perception for Improving Face Recognition Under Occlusions

摘要

Recent studies have demonstrated that active perception can improve the perception abilities of deep learning (DL) models. However, there are challenges associated with using active perception in DL models, including the need for datasets and/or realistic simulations that can support the training process, along with the difficulty of predicting the final target position, which reduces planning efficiency. To address these challenges, this work presents a methodology for enhancing the perception abilities of DL models through active perception. The methodology proposes a way to create datasets for active perception by fusing existing large-scale datasets and decomposing the active perception problem into three subtasks for face recognition. The subtasks aim to determine the appropriateness of the current view for face recognition, the direction in which the robot should move for a better viewpoint, and the expected amount of movement required. A novel trial-based approach is introduced to estimate the final target position, making the method platform-agnostic and easily applicable to different robots. The proposed methodology is validated through experiments on two well-known face verification datasets that have been augmented with occlusions, demonstrating its effectiveness in enhancing the perception abilities of DL models through active perception.