EDMB: Edge Detector with Mamba

TL;DR

This work introduces EDMB, a Mamba-based edge detector that combines a global Mamba encoder, a fine-grained Mamba encoder, and a lightweight high-resolution CNN branch to efficiently model global and local context for edge discovery. A Learnable Gaussian Distributions Decoder generates multi-granularity edges by sampling from learned distributions, with training guided by an Evidence Lower Bound (ELBO) loss that couples a KL term and a weighted cross-entropy. Key contributions include the multi-stage training strategy, a CFF-based fusion for means and variances, and the ability to produce multi-granularity edges on single-label datasets via ELBO supervision. Empirically, EDMB achieves competitive results on BSDS500 (ODS $0.837$ single, $0.851$ multi) and strong performance on NYUDv2 and BIPED, while offering efficiency advantages over Transformer-based methods. This approach advances edge detection by delivering high-quality edges with flexible granularity and reduced reliance on multi-label annotations.

Abstract

Transformer-based models have made significant progress in edge detection, but their high computational cost is prohibitive. Recently, vision Mamba have shown excellent ability in efficiently capturing long-range dependencies. Drawing inspiration from this, we propose a novel edge detector with Mamba, termed EDMB, to efficiently generate high-quality multi-granularity edges. In EDMB, Mamba is combined with a global-local architecture, therefore it can focus on both global information and fine-grained cues. The fine-grained cues play a crucial role in edge detection, but are usually ignored by ordinary Mamba. We design a novel decoder to construct learnable Gaussian distributions by fusing global features and fine-grained features. And the multi-grained edges are generated by sampling from the distributions. In order to make multi-granularity edges applicable to single-label data, we introduce Evidence Lower Bound loss to supervise the learning of the distributions. On the multi-label dataset BSDS500, our proposed EDMB achieves competitive single-granularity ODS 0.837 and multi-granularity ODS 0.851 without multi-scale test or extra PASCAL-VOC data. Remarkably, EDMB can be extended to single-label datasets such as NYUDv2 and BIPED. The source code is available at https://github.com/Li-yachuan/EDMB.

Figures (4)

• Figure 1: The EDMB's framework. $N(\mu,\sigma^2)$ means learnable Gaussian distributions, where the means $\mu$ and variances $\sigma^2$ are predicted by the mean decoder and variance decoder, respectively.
• Figure 2: The detailed architecture of the LGD decoder. $MBConv$ is a kind of efficient convolution proposed by MobileNetv2 howard2017mobilenets.
• Figure 3: Qualitative comparisons on challenging samples in the BSDS500 test set. MuGE produces diverse results with edge granularity of 0, 0.5, and 1, respectively. EDMB and UAED samples from the learned distribution with $\mu+\gamma\sigma^2$. Fig. (c)$\sim$(l) come from zhou2024muge.
• Figure 4: Results of different multi-label handling methods. The first row shows the results for the mixed label, and the second row shows the results for the randomly selected label