DGNet

DualGazeNet: A Biologically Inspired Dual-Gaze Query Network for Salient Object Detection

Yu Zhang, Haoan Ping, Yuchen Li, Zhenshan Bing, Wei He, Fellow, IEEE, Fuchun Sun, Fellow, IEEE, Alois Knoll, Fellow, IEEE

Technical University of Munich

Paper:(arxiv:2511.18865)

Abstract

DualGazeNet is a biologically inspired Transformer framework for salient object detection, designed with dual-path processing inspired by the human visual system. It achieves state-of-the-art performance on five RGB SOD benchmarks as well as 4 COD benchmarks and USOD10K dataset.

Overview

News

• Nov 25, 2025: We released our paper on arxiv.

• Nov 20, 2025: We released the well-trained weights under different configs for SOD/COD/USOD tasks with various resolutions. We also provide the corresponding datasets, the pretrained backbone weights and their prediction maps from both our models and other SOTA models.

• Nov 17, 2025: We released DGNet codes.

Usage

Installation

Step 1:

Clone this repository

git clone https://github.com/jeremypha/DualGazeNet.git
cd DualGazeNet

Step 2:

Create a new conda environment

conda create --name dgnet
conda activate dgnet

Install Dependencies

pip install torch --index-url https://download.pytorch.org/whl/cu<your cuda version>
pip install -r requirements.txt

Set up Datasets

-- datasets
    |-- SOD
    |    |-- DUTS-TR
    |    |   |-- im
    |    |   |-- gt
    |    |-- DUTS-TE
    |    |   |-- im
    |    |   |-- gt
    |    |-- DUT-OMRON
    |    |   |-- im
    |    |   |-- gt
    |    |-- ECSSD
    |    |   |-- im
    |    |   |-- gt
    |    |-- HKU-IS
    |    |   |-- im
    |    |   |-- gt
    |    |-- PASCAL-S
    |    |   |-- im
    |    |   |-- gt
    |-- COD
    |-- USOD

All datasets are publicly available from their official sources: DUTS, DUT-OMRON, HKU-IS, ECSSD, and PASCAL-S.

For convenience, we provide pre-configured versions with consistent formatting in our BaiduNetDisk Folder, which also includes datasets for COD and USOD tasks.

Train

Download Pretrained Backbones and save it in ./weights

./scripts/train.sh TASK BACKBONE

# Example: ./scripts/train.sh SOD L

Evaluation and Predicted Saliency Map

Model weights and corresponding prediction maps for all configurations are available for download. Access the full dataset in our Google Drive Folder, or retrieve specific items individually from the following table.

Task	Backbone	Resolution	Params(M)	FLOPs(G)	FPS	Checkpoint	Saliency Map
SOD	Hiera-L	512×512	247.56	238.52	43	checkpoint	Results
SOD	Hiera-L	352×352	247.56	139.07	45	checkpoint	Results
SOD	Hiera-L	224×224	247.56	48.59	46	checkpoint	Results
SOD	Hiera*-L	512×512	162.32	217.11	48	checkpoint	Results
SOD	Hiera*-L	352×352	162.32	126.27	50	checkpoint	Results
SOD	Hiera*-L	224×224	162.32	44.19	52	checkpoint	Results
SOD	Hiera-B	512×512	91.92	102.78	61	checkpoint	Results
SOD	Hiera-B	352×352	91.92	47.95	64	checkpoint	Results
SOD	Hiera-B	224×224	91.92	17.89	69	checkpoint	Results
SOD	Hiera*-B	512×512	49.23	83.47	72	checkpoint	Results
SOD	Hiera*-B	352×352	49.23	39.13	77	checkpoint	Results
SOD	Hiera*-B	224×224	49.23	18.86	78	checkpoint	Results
COD	Hiera-L	512×512	247.56	238.52	43	checkpoint	Results
USOD	Hiera-L	512×512	247.56	238.52	43	checkpoint	Results

./scripts/inference.sh TASK BACKBONE CHECKPOINT

# Example: ./scripts/inference.sh SOD L ./output/epoch_0.pth

Quantitative Comparison

Here you can download saliency maps of SOD/COD/USOD tasks from other awesome models: BaiduNetDisk

Citation