Research map / Detection and segmentation / CNN features
Frequency-Adaptive Dilated Convolution for Semantic Segmentation
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 Frequency-Adaptive Dilated Convolution (FADC), combining adaptive dilation rates, frequency-decomposed adaptive kernels, and frequency selection to improve dilated convolution for semantic segmentation and detection.
Problem
Standard dilated convolution uses a fixed global dilation rate, which shrinks effective bandwidth (causing gridding/aliasing artifacts) and limits receptive field, creating a trade-off that is poorly balanced across spatially varying features.
Approach
Analyzes dilated convolution via spectrum analysis and makes the dilation rate spatially adaptive (AdaDR) predicted from local high-frequency power, adapts per-channel low/high-frequency kernel weights (AdaKern), and reweights frequency bands spatially in the input (FreqSelect) to balance bandwidth and receptive field.
Key results
- On Cityscapes, PIDNet-M-FADC reaches 81.0 mIoU at 37.7 FPS, surpassing the heavier PIDNet-L (80.9 val mIoU at 31.1 FPS).
- On ADE20K with UPerNet, ResNet-50-FADC improves mIoU from 40.7 to 44.4 (SS), above ResNet-101 at 42.9.
- Theoretical average receptive field of dilated ResNet-50 rises from 441 to 1100 pixels with FADC plus FreqSelect.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace dilated convolution with FADC | PSPNet with Dilated-ResNet-50 · Cityscapes val | mIoU: 77.8 -> 80.4 | Helped |
| Replace DCNv2 with FADC | PSPNet with Dilated-ResNet-50 · Cityscapes val | mIoU: 79.7 -> 80.4 | Helped |
| Replace dilated convolution with FADC | DeepLabV3+ with Dilated-ResNet-50 · Cityscapes val | mIoU: 79.2 -> 80.3 | Helped |
| Replace DCNv2 with FADC | DeepLabV3+ with Dilated-ResNet-50 · Cityscapes val | mIoU: 79.9 -> 80.3 | Helped |
| Replace ADC with FADC | DeepLabV3+ with Dilated-ResNet-101 · Cityscapes val | mIoU: 80.7 -> 81.5 | Helped |
| Replace convolution with FADC | Mask2Former with ResNet-50 · Cityscapes val | mIoU: 79.4 -> 80.6 | Helped |
| Replace convolution with FADC | ResNet-50 with UPerNet · ADE20K val | mIoU (SS): +3.7 | Helped |
| Add FADC | PIDNet-M · Cityscapes val | mIoU: 80.1 -> 81.0 | Helped |
| Add FreqSelect | InternImage-T · ADE20K val | mIoU (SS): 47.9 -> 48.7 | Helped |
| Add FreqSelect | DiNAT-S · COCO | APmask: +0.6 | Helped |
| Add AdaKern and FreqSelect | ResNet-50 + DCNv2 with Faster R-CNN · COCO, 1× | APbox: 41.3 -> 42.2 | Helped |
| Add FreqSelect (ablation) | FADC · Dilated ResNet-50 theoretical receptive field | Receptive field: 1007 -> 1100 | Helped |
Limitations
- Extension to deformable and dilated attention is left as future work; the frequency analysis is only applied to convolution.
- Spatial dilation prediction and FreqSelect add some computation and parameters compared with plain convolution.
Experimental setup
Datasets: Cityscapes ADE20K COCO
Benchmarks: Cityscapes ADE20K COCO
Metrics: mIoU AP FPS FLOPs receptive field
Baselines: Dilated Convolution DCNv2 ADC (Adaptive Dilated Convolution) ConvNeXt-B-dcls PIDNet-L InternImage-T DiNAT-S
Builds on: dilated convolution deformable convolution v2 InternImage DiNAT Mask2Former
Closest papers by meaning
- Frequency Dynamic Convolution for Dense Image Prediction (CVPR 2025 · similarity 0.89)
- When Semantic Segmentation Meets Frequency Aliasing (ICLR 2024 · similarity 0.86)
- Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.86)
- FFR: Frequency Feature Rectification for Weakly Supervised Semantic Segmentation (CVPR 2025 · similarity 0.86)
- Multiscale Low-Frequency Memory Network for Improved Feature Extraction in Convolutional Neural Networks (AAAI 2024 · similarity 0.84)
- Learning Frequency-Adapted Vision Foundation Model for Domain Generalized Semantic Segmentation (NeurIPS 2024 · similarity 0.84)
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.