LISA: Reasoning Segmentation via Large Language Model
Keywords: Large Language Model, Multi-model Large Lanaguage Model, Segmentation, Reasoning Segmentation
Abstract: Although perception systems have made remarkable advancements in recent years, they still rely on explicit human instruction to identify the target objects or categories before executing visual recognition tasks. Such systems lack the ability to actively reason and comprehend implicit user intentions. In this work, we propose a new segmentation task --- reasoning segmentation. The task is designed to output a segmentation mask given a complex and implicit query text. Furthermore, we establish a benchmark comprising over one thousand image-instruction pairs, incorporating intricate reasoning and world knowledge for evaluation purposes. Finally, we present LISA: large Language Instructed Segmentation Assistant, which inherits the language generation capabilities of the multi-modal Large Language Model (LLM) while also possessing the ability to produce segmentation masks. We expand the original vocabulary with a <SEG> token and propose the embedding-as-mask paradigm to unlock the segmentation capability. Remarkably, LISA can handle cases involving: 1) complex reasoning; 2) world knowledge; 3) explanatory answers; 4) multi-turn conversation. Also, it demonstrates robust zero-shot capability when trained exclusively on reasoning-free datasets. In addition, fine-tuning the model with merely 239 reasoning segmentation image-instruction pairs results in further performance enhancement. Experiments show our method not only unlocks new reasoning segmentation capabilities but also proves effective in both complex reasoning segmentation and standard referring segmentation tasks.
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/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: 1126
Loading