Zero-shot Human-Object Interaction Detection via Conditional Multi-Modal Prompts
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 Object Interaction Detection, Zero-shot
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Human Object Interaction (HOI) detection is the task of locating and inferring the relationships between all possible human-object combinations. One of the most challenging issues is the extensive labor required for the annotation of combinatorial space of possible HOI interactions. Most existing HOI detectors rely on full annotations of all predefined interactions, resulting in a lack of generalisation for unseen combinations and actions. Inspired by the powerful generalisation ability of the large Vision-Language Models (VLM), we propose a Prompt-based zero-shot human-object Interaction Detection framework, namely PID, which can improve alignment between the vision and language representations using conditional multi-modal prompts. Specifically, different from traditional prompt-learning methods, we propose learning decoupled visual and language prompts for spatial-aware visual feature extraction and interaction classification, respectively. Furthermore, we introduce constraints for multi-modal prompts to alleviate the problem of overfitting to seen concepts in prompt learning process, thus improving the suitability for zero-shot settings. Extensive experiments demonstrate the prominence of our detector with conditional multi-modal prompts, outperforming previous state-of-the-art on unseen classes of various zero-shot settings.
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: 8729
Loading