GTD-LLM: A Plug-and-Play LLM Reasoning Module for Gaze Target Detection
Keywords: LLM Reasoning; Prompt Engineering; Gaze Target Detection
TL;DR: We propose GTD-LLM, the first plug-and-play module which leverages the powerful logical reasoning ability of LLMs to address gaze target detection in visual scenes.
Abstract: Gaze target detection is an important task in computer vision, aiming to predict where people in an image are looking. In our view, this task not only contains explicit image features, but also implies a large amount of prior knowledge about the correlations between human visual attention and daily activities. However, existing gaze target methods rely entirely on visual modality information to detect salient objects along the gaze direction, limiting their generalization in challenging scenarios such as activity-related, long-tailed, small-sized, or long-distance gaze targets. Inspired by the great success of LLM technology, we break away from the traditional pure-visual approaches and propose GTD-LLM, the first plug-and-play LLM reasoning module for gaze target detection in visual scenes, providing a new paradigm for traditional pure-visual approaches. Our GTD-LLM module can be plug-and-play integrated with any existing gaze target visual models and directly bring them universal performance improvements, simultaneously demonstrating strong generalizability and effectiveness. In our GTD-LLM module, we design a novel prompt engineering method GTD-Prompt, to guide LLMs like GPT-4 to perform logical reasoning on possible gaze targets, without the need for any training or fine-tuning. The proposed GTD-Prompt method can also be easily extended to downstream tasks by simply adjusting the corresponding task prompt words, further illustrating its versatility.
Primary Area: applications to 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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 1738
Loading