Parameter-Efficient Domain Adaptation for Edge Detectors: A Comparative Study of Colormap Models and LoRA

Abstract Edge Detectors are one of the most fundamental tools in Computer Vision applications. Using deep learning, state-of-the-art (SOTA) models can 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 limited data severely hinders performance on images from other domain spaces. This paper proposes 2 methods to adapt SOTA models to single-channel images with minimal computational overhead. We first propose a feed-forward network (FFN) for domain adaptation. Then, we propose an extension on this network that uses Fourier convolutions to boost performance. Our models serve as an additional layer to existing SOTA edge detectors, allowing the edge detector to generalize well to new domain spaces without undergoing extensive resource or data-heavy retraining. Our models were trained with a single off-the-shelf GPU with trainable parameters less than 60K and a small dataset of 30-50 images. The training of the FFN took 30 minutes and the training of the Fourier convolution model took 2 hours. Using these lightweight models, we could boost the F-score of the resulting edgemaps by over 0.2.