Segmenting the Unknown: Discrete Diffusion Models for Non-Deterministic Segmentation
Primary Area: generative models
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: segmentation, diffusion, future-prediction
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Discrete diffusion models to handle ambiguity and uncertainty in semantic segmentation of medical images and future driving scene.
Abstract: Safety critical applications of deep-learning require models able to handle ambiguity and uncertainty.
We introduce discrete diffusion models to capture uncertainty in semantic segmentation, with application in both oncology and autonomous driving.
Unlike prior approaches that tackle these tasks in distinct ways, we formulate both as estimating a complex posterior distribution over images, and present a unified solution that leverages the discrete diffusion framework.
Our contributions include the adaptation of discrete diffusion for semantic segmentation to model uncertainty and the introduction of an auto-regressive diffusion framework for future forecasting.
Experimental evaluation on medical imaging data and real-world future prediction tasks demonstrates the superiority of our generative framework over deterministic models and its competitive performance compared to methods specific to these domains separately.
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: 5672
Loading