홈 › Papers

A Doubly Decoupled Network for edge detection

Deep learning-based edge detectors are increasingly scaled up to achieve higher performance. However, this growth comes at the cost of higher computational demands and a greater risk of model overfitting. In this paper, a Doubly Decoupled Network (DDN) is proposed to alleviate these problems by decoupling the data and features separately. Firstly, we decouple image-edge pairs and build learnable Gaussian distributions. The images are mapped to learnable Gaussian distributions, and the corresponding edges are mapped to samples of the distributions. More robust features are generated by normalizing the latent variables and sampling from the Gaussian distributions. Secondly, we compress the redundant calculations by decoupling image features. Shallow features need not only high resolution to provide accurate spatial information, but also diversity to provide sufficient cues for deep features. We decouple image features into spatial and semantic features and encode them separately. Shallow features no longer provide cues for semantic features and their diversity can be drastically compressed. The proposed DDN achieves the state-of-the-art accuracy across multiple edge detection benchmarks, while the computational cost is similar to that of VGG-based methods. The code is available at https://github.com/Li-yachuan/DDN.

Code (1)

Tasks

Similar Papers 제목 키워드 기반