Research map / Detection and segmentation / CNN features
ESNet: Evolution and Succession Network for High-Resolution Salient Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ESNet, a two-stage high-resolution salient object detection model with a low-resolution evolution stage (supervision and feature evolution) and a lightweight high-resolution succession stage, plus a new MAEBD metric.
Problem
High-resolution salient object detection (HRSOD) must preserve fine details while avoiding the high computational cost of processing full-resolution images.
Approach
Decouples HRSOD into low-resolution saliency localization with evolutionary supervision (progressively dilated ground truth per decoder layer) and feature-level evolutionary variation constraints, then performs lightweight high-resolution refinement using shallow features, prototype-based boundary enhancement (SFEM), and mutually exclusive supervision, without patch cropping.
Key results
- On HRSOD-TE, ESNet reaches F-measure 0.926 and MAEBD 0.166 versus HQSOD's 0.907 and 0.185 (MAE 0.021 vs 0.024).
- Runs at 49 FPS on 1280x1280 inputs on a single NVIDIA 3090, versus under 5 FPS for HQSOD at 1024x1024.
- Removing the evolution mechanisms on UHRSD-TE raises MAEBD from 0.164 to 0.176 and lowers F-measure from 0.915 to 0.905.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the succession-stage high-resolution refinement to LrLM (ablation) | ESNet · HRSOD-TE | MAEBD: 0.206 -> 0.166 | Helped |
| add patch-dividing in the succession stage (ablation) | ESNet · HRSOD-TE | MAEBD: 0.166 -> 0.190 | Hurt |
| add evolution on supervision and feature (ablation) | ESNet · UHRSD-TE | F-measure: 0.905 -> 0.915 | Helped |
| add feature evolution (ablation) | ESNet with supervision evolution · UHRSD-TE | F-measure: 0.906 to 0.911 | Mixed |
| add the evolutionary variation loss (ablation) | ESNet · HRSOD-TE | MAEBD: 0.168 -> 0.166 | Helped |
| add mutually exclusive supervision (MES) (ablation) | ESNet · HRSOD-TE | MAEBD: 0.171 -> 0.166 | Helped |
| restrict prototype extraction and feature enhancement to contour regions instead of whole foreground and background regions (ablation) | ESNet SFEM · HRSOD-TE | MAEBD: 0.180 -> 0.166 | Helped |
| use SFEM instead of standard self-attention (ablation) | ESNet · HRSOD-TE | MAEBD: 0.201 -> 0.166 | Helped |
| use ESNet rather than PGNet | HRSOD-TE · HRSOD-TE | F-measure: 0.928 -> 0.937 | Helped |
| use ESNet rather than PGNet | DAVIS-SOD · DAVIS-SOD | MAE: 0.012 -> 0.009 | Helped |
| use ESNet rather than PGNet | UHRSD-TE · UHRSD-TE | MAEBD: 0.155 -> 0.147 | Helped |
Surprising findings
- Adding only supervision evolution without feature evolution gave no overall improvement on HRSOD, which the authors attribute to the network not learning the relationship between label levels.
Limitations
- Evaluation limited to three high-resolution datasets and a CNN (ResNet50) and Swin variant.
- MAEBD depends on hand-set patch size (80x80) and perimeter-to-area threshold (0.25).
Experimental setup
Datasets: HRSOD DAVIS-SOD UHRSD
Benchmarks: HRSOD-TE DAVIS-SOD UHRSD-TE
Metrics: MAE F-measure S-measure FNR MAEBD
Baselines: BASNet MINet GateNet GCPANet PFSNet MSFNet CTDNet ICON EDN HRSOD DRFNet HQSOD
Builds on: CTDNet ResNet PGNet HQSOD
Closest papers by meaning
- Salient Object Detection Driven by Fixation Prediction (CVPR 2018 · similarity 0.88)
- BASNet: Boundary-Aware Salient Object Detection (CVPR 2019 · similarity 0.87)
- Locate Globally, Segment Locally: A Progressive Architecture With Knowledge Review Network for Salient Object Detection (AAAI 2021 · similarity 0.87)
- An Iterative and Cooperative Top-Down and Bottom-Up Inference Network for Salient Object Detection (CVPR 2019 · similarity 0.86)
- Detect Globally, Refine Locally: A Novel Approach to Saliency Detection (CVPR 2018 · similarity 0.86)
- No Pains, More Gains: Recycling Sub-Salient Patches for Efficient High-Resolution Image Recognition (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.