EDMB: Edge Detector with Mamba

May 29, 2026 · View on GitHub

Update

We released "A Doubly Decoupled Network for Edge Detection"(DDN)

Main features: Transformer is used as the main feature extraction module and one-stage training is adopted, which is more convenient than EDMB. It uses the same multi-granularity edge generation strategy as EDMB. The accuracy is better than EDMB, and the ODS on BSDS500 reaches 0.867.

Update 260416

Uploaded images for multi-granularity testing on the BSDS500 (BSDS_MG.zip), NYUD (NYUD_MG1.zip, NYUD_MG2.zip), and BIPED (BIPED_MG.zip) datasets. Please note that the NYUD data was split into two parts due to file size limits; please merge them before use.

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 depen- dencies. 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.

The EDMB’s framework.

Fig. 1 The EDMB’s framework. N(µ, σ2) means learnable Gaussian distributions, where the means µ and variances σ2 are predicted by the mean decoder and variance decoder, respectively.

Result on bsds.

Table 1 Quantitative results on the BSDS500 dataset. All edges are generated without multi-scale testing or additional PASCAL-VOC data. * means generating multi-granularity edges. The best two results are denoted as red and blue respectively, and the same for other tables.

Vis on bsds.

Fig. 2 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 µ + γσ^2

For more details, see the paper

Getting Started

Step 1: Clone the EDMB repository:

git clone https://github.com/Li-yachuan/EDMB.git
cd EDMB

setup virture env

conda create -n edmb
conda activate edmb

pip install -r requirements.txt

Install Dependencies for vmamba

wget https://github.com/Dao-AILab/causal-conv1d/archive/refs/tags/v1.1.1.zip
unzip v1.1.1.zip
cd causal-conv1d-1.1.1 && pip install .
wget https://github.com/state-spaces/mamba/archive/refs/tags/v1.1.1.zip
unzip v1.1.1.zip
cd mamba-1.1.1 && pip install .
wget https://github.com/MzeroMiko/VMamba/tree/main/kernels/selective_scan
cd kernels/selective_scan && pip install .

Step 3: download data and pretrained-parameter

dataset

BSDS500: Following UAED&MuGE. For better visualization, we convert the images in mat format to jpg in advance. If you use images in mat format, you need to make a simple change in the data loading module.
NYUDv2: Following RCF.
BIPED: Following DexiNed. download them and point to the path in main.py(start from line 73)

download ckpt

vmabma-s and put it to ./model/

cd model && wget https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_small_0229_ckpt_epoch_222.pth

Step 4: training model

stage I

python main.py --batch_size 4 --stepsize 10-16 --maxepoch 20 --gpu 1 --encoder DUL-Mamba-s --savedir [save dir] --dataset BSDS-rand

stage II

python main.py --batch_size 3 --stepsize 10-14 --maxepoch 16 --gpu 2 --encoder MIXENC_PNG --decoder MIXUNET --savedir [save dir] --dataset BSDS-rand --global_ckpt [bset result of Stage I]

Generating multi-granu edge

python main.py --batch_size 3 --stepsize 10-14 --maxepoch 16 --gpu 2 --encoder MIXENC_PNG --decoder MIXUNET --savedir [save dir] --dataset BSDS-rand --global_ckpt [bset result of Stage I] --mode test --resume [bset result of Stage II] -mg

python eval_muge_best/best_ods_ois_full.py [result dir]

For multi-granularity testing, refer to MuGE

Acknowledgments

UAED MuGE EDTER VMAMBA