Embodied Referring Expression Comprehension Through Multimodal Residual Learning

Md Mofijul Islam, Alexi Gladstone, Sujan Sarker, Ganesh Nanduru, Md Fahim, Keyan Du, Aman Chadha, Tariq Iqbal

Keywords: embodied, multimodal, visual-language

TL;DR: We introduce Refer360, a dataset of real-world human interactions, and MuRes, a multimodal residual learning module that improves models' comprehension of verbal and nonverbal cues, outperforming existing systems in complex environments.

Abstract: Comprehending embodied interactions within real-world settings poses a considerable challenge, attributed to the multifaceted nature of human interactions and the variability of environments, necessitating the development of comprehensive benchmark datasets and multimodal learning models. Existing datasets do not adequately represent the full spectrum of human interactions, are limited by perspective bias, rely on single viewpoints, have insufficient nonverbal gesture capture, and have a predominant focus on indoor settings. To address these gaps, we present an Embodied Referring Expressions dataset (called Refer360), which contains an extensive collection of embodied verbal and nonverbal interaction data captured from various viewpoints across various indoor and outdoor settings. In conjunction with this benchmark dataset, we propose a novel multimodal guided residual module (MuRes) that helps the existing multimodal models to improve their representations. This guided residual module acts as an information bottleneck to extract salient modality-specific representations, and reinforcing these to the pre-trained representations produces robust complementary representations for downstream tasks. Our extensive experimental analysis of our benchmark Refer360 dataset reveals that existing multimodal models alone fail to capture human interactions in real-world scenarios comprehensively for embodied referring expression comprehension tasks. Building on these findings, a thorough analysis of four benchmark datasets demonstrates superior performance by augmenting MuRes into current multimodal models, highlighting its capability to improve the understanding and interaction with human-centric environments. This paper offers a benchmark for the research community and marks a stride towards developing robust systems adept at navigating the complexities of real-world human interactions.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11352