Selective LoRA for Domain-Aligned Dataset Generation in Urban-Scene Segmentation

22 Sept 2024 (modified: 05 Feb 2025)Submitted to ICLR 2025EveryoneRevisionsBibTeXCC BY 4.0
Keywords: Dataset Generation, Urban-scene Segmentation
TL;DR: We propose Selective LoRA, a novel fine-tuning method that generates well-aligned and informative segmentation datasets by updating only weights related to desired concepts. We improve existing urban-scene segmentation models in various settings.
Abstract:

This paper addresses the challenge of data scarcity in semantic segmentation by generating datasets through fine-tuned text-to-image generation models, reducing the costs of image acquisition and labeling. Segmentation dataset generation faces two key challenges: 1) aligning generated samples with the target domain and 2) producing informative samples beyond the training data. Existing methods often overfit and memorize training data, limiting their ability to generate diverse and well-aligned samples. To overcome these issues, we propose Selective LoRA, a novel fine-tuning approach that selectively identifies and updates only the weights associated with necessary concepts (e.g., style or viewpoint) for domain alignment while leveraging the pretrained knowledge of the image generation model to produce more informative samples. Our approach ensures effective domain alignment and enhances sample diversity. We demonstrate its effectiveness in generating datasets for urban-scene segmentation, outperforming baseline and state-of-the-art methods in in-domain (few-shot and fully-supervised) settings, as well as domain generalization tasks, especially under challenging conditions such as adverse weather and varying illumination, further highlighting its superiority.

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.
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: 2521
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview