Learning Attribute Attention and Retrospect Location for Instance Object Navigation

Abstract

Listen

Translate article

Visual object navigation, a classical problem in embodied intelligence tasks, requires agents to find a specified target using only the first-person view of visual information. A number of methods that search only for an object in the same category are known as category navigation. This is different from the real world where objects are labelled as “whose”. Then, the navigation task of distinguishing objects by using attributes is closer to practicality, which is called instance-level object navigation. However, the current instance-level object navigation approaches have some limitations. Firstly, different attribute information are jumbled together and the fine-grained feature relationships are ignored, resulting weakened discriminative ability. Secondly, the explored trajectories have insufficient correlation with the navigation target, resulting in degraded memory ability. In this paper, we present a novel cascade architecture from a fresh perspective to solve those limitations. Our approach has two main techniques: Object-Attribute Attention Graph (OAAG) and Objective Retrospect and Location Module (ORLM). OAAG consists of Object-aware Graph (OAG) which is created to encode the dynamically relations between all instances, and Attribute-Attention Graph (AAG) which assigns different attention weight to different attributes. ORLM makes it possible for the agent to review the region it explores and enhance the relevant memorization of the target. We connect them as the final model output and input it to the deep reinforcement learning A3C framework. We evaluate these two technologies on the AI2-THOR simulator. Experimental studies have shown that our approach outperforms other related works and achieves the state-of-the-art results on three specific tasks (Instance-Localization, Instance-Navigation and Category-Localization). The project is available at https://github.com/visee-sdu/Instance-Navigation.