Sub-Domain Aware Granular Segmentation via Fine Tuning Network
Keywords: domain adaptation
Abstract: Recent advances in deep learning (DL) have led to improved vision-based algorithms. DL-based semantic segmentation, in particular, has enabled precise predictions using Convolutional Neural Networks (CNNs). State-of-the-art CNN-based networks have achieved high accuracy on various datasets in multiple fields, such as building, scene, and object segmentation. However, subdomain shifts between training and test sets within a single domain can cause degraded accuracy in fine-grained segmentation. To counter this, this paper introduces a novel Sub-Domain Adaptation (SDA) framework for fine-grained and granular segmentation, which divides one single domain into multiple sub-domains and optimizes the baseline-network for each sub-domain. The baseline-network is further fine-tuned by recognizing the domain of the input in run-time, leading to more accurate predictions. Benchmarks of scene parsing, autonomous driving, and aerial imagery demonstrate the superior performance of SDA for granular segmentation.
Primary Area: optimization
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: 11314
Loading