Title:SeekNet: Improved Human Instance Segmentation via Reinforcement Learning Based Optimized Robot Relocation

Abstract:Amodal recognition is the ability of the system to detect occluded objects. Most state-of-the-art Visual Recognition systems lack the ability to perform amodal recognition. Few studies have achieved amodal recognition through passive prediction or embodied recognition approaches. However, these approaches suffer from challenges in real-world applications, such as dynamic objects. We propose SeekNet, an improved optimization method for amodal recognition through embodied visual recognition. Additionally, we implement SeekNet for social robots, where there are multiple interactions with crowded humans. Hence, we focus on occluded human detection & tracking and showcase the superiority of our algorithm over other baselines. We also experiment with SeekNet to improve the confidence of COVID-19 symptoms pre-screening algorithms using our efficient embodied recognition system.