Domain Adaptation for Edge Detectors Using Lightweight Networks
Venkat Sumanth Reddy Bommireddy, Sophie Raniwala, Piyush Kumar · 2023
Edge Detectors are one of the most fundamental tools used in Computer Vision applications. Using deep learning, state-of-the-art (SOTA) models are able to produce sharp, fine edges, mirroring human-level performance. However, most of these SOTA models are trained solely on color images. We experimentally determined that this skewed data severely hinders performance on images from other domain spaces. In this paper, we propose a way to adapt SOTA models to single-channel inputs from other domains with minimal overhead, using small-scale feed-forward networks. Our model, which serves as an additional layer to existing SOTA models, greatly boosts their performance on out-of-domain data, allowing the edge detector to generalize to new domain spaces without undergoing extensive, resource-or data-heavy retraining. Using a single off-the-shelf GPU and a small 30-image dataset, we were able to train this low-complexity (less than 13k parameters) model in half an hour to boost the F-score of the produced edgemaps for chest x-rays by over 0.2.