Sparse Mask Representation for Human-Scene Interaction

Anh Nguyen, Nghia Huu Vu, Tuong Khanh Long Do, Le Duc Huy, Erman Tjiputra, Quang D. Tran

Primary Area: representation learning for computer vision, audio, language, and other modalities

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.

Keywords: Human-scene Interaction, Scene Synthesis

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Human-scene interaction is an active research topic with several applications in robotics, virtual experiences, gaming, surveillance, and healthcare. Despite efforts to improve the network architectures to achieve better results or optimize models for faster inference, a crucial aspect of input dimensionality has been somewhat overlooked. This paper introduces Sparse Mask Representation, a simple yet effective approach to enhance the inference speed of human-scene interaction models and improve the model's effectiveness by exploring the sparsity of high-dimensional inputs. Specifically, our method utilizes sparse masks to convert high-dimensional inputs into sparse tensors in a compressed COO format. Our approach not only effectively streamlines computational speed but also eliminates non-useful input information, thereby enhancing overall model performance. We conducted rigorous experiments across three datasets, with a specific emphasis on tasks related to contact prediction and scene synthesis. The results underscore the substantial enhancements realized by our proposed method in terms of accuracy and inference time, surpassing existing state-of-the-art approaches.

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: 1581