Research map / Detection and segmentation / CNN features
FD2-Net: Frequency-Driven Feature Decomposition Network for Infrared-Visible 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 FD2-Net, a YOLOv5-based infrared-visible object detector that splits features into DCT-based high-frequency and multi-dilation low-frequency branches, recouples them parameter-free, and adds a mask-based cross-reconstruction module, reporting SOTA mAP50 on LLVIP, FLIR and M3FD.
Problem
Infrared-visible object detection often ignores the frequency characteristics of complementary information (high-frequency visible detail, low-frequency infrared thermal content), uses fixed receptive fields, and loses object detail through repeated downsampling.
Approach
A feature decomposition encoder with a DCT-based high-frequency unit and a dilated multi-kernel low-frequency unit, followed by a parameter-free cross-modal complementary-strengths addition and a cross-reconstruction unit with asymmetric feature masking that reconstructs the input images to restore lost detail.
Key results
- FD2-Net reaches 96.2% mAP50 on LLVIP, versus 95.4% for the next best (TFDet).
- Achieves 82.9% mAP50 and 42.5% mAP75 on FLIR, improving on the prior best by about 1.5 points at mAP75.
- Reaches 83.5% mAP50 on M3FD, outperforming SegMiF (82.2%) and CDDFuse (81.9%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FD2-Net instead of TFDet | YOLOv5-based IVOD on LLVIP · LLVIP | mAP50: 95.4 -> 96.2 | Helped |
| Use FD2-Net instead of TFDet | YOLOv5-based IVOD on FLIR · FLIR | mAP50: 81.7 -> 82.9 | Helped |
| Use FD2-Net instead of SegMiF | YOLOv5-based IVOD on M3FD · M3FD validation | mAP50: 82.2 -> 83.5 | Helped |
| Add HFU (ablation) | FD2-Net component ablation on LLVIP · LLVIP | mAP50: 90.9 -> 92.4 | Helped |
| Add LFU alongside HFU (ablation) | FD2-Net component ablation on LLVIP · LLVIP | mAP50: 92.4 -> 94.7 | Helped |
| Add CRU (ablation) | FD2-Net component ablation on LLVIP · LLVIP | mAP50: 94.7 -> 95.6 | Helped |
| Add complementary mask (ablation) | FD2-Net component ablation on LLVIP · LLVIP | mAP50: 95.6 -> 96.2 | Helped |
| Replace C2f blocks with HFU only (ablation) | YOLOv5n · LLVIP | mAP50: 90.9 % -> 91.5 % | Helped |
| Replace C2f blocks with LFU only (ablation) | YOLOv5n · LLVIP | mAP50: 90.9 % -> 91.9 % | Helped |
| Use parallel HFU and LFU instead of sequential HFU then LFU (ablation) | YOLOv5n · LLVIP | mAP50: 92.8 % -> 94.7 % | Helped |
| Use parallel HFU and LFU instead of sequential LFU then HFU (ablation) | YOLOv5n · LLVIP | mAP50: 92.6 % -> 94.7 % | Helped |
Surprising findings
- Using HFU or LFU alone in place of C2f blocks reduced mAP50 relative to the YOLOv5n baseline (drops of 3.2% and 2.8%, as reported in the text), while the parallel combination gave the best result.
Limitations
- Gains over prior methods are modest on some metrics and no variance or multiple-seed results are reported.
- The paper notes that more complex attention architectures could potentially yield further improvement but were not explored.
- Asymmetric mask ratio and loss weights are described but their sensitivity is not analyzed in the provided text.
Experimental setup
Datasets: LLVIP FLIR M3FD
Benchmarks: LLVIP FLIR M3FD
Metrics: mAP50 mAP75 F1 Precision Recall
Baselines: Infrared Visible DensFuse SDNet U2Fusion CDDFuse MetaF LRRNet SegMiF TarDAL DDFM CSSA
Builds on: YOLOv5 ConvNeXt RepLKNet
Closest papers by meaning
- HS-FPN: High Frequency and Spatial Perception FPN for Tiny Object Detection (AAAI 2025 · similarity 0.86)
- FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network (NeurIPS 2025 · similarity 0.85)
- Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm (NeurIPS 2026 · similarity 0.84)
- IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · similarity 0.84)
- Fine-Grained Prototypes Distillation for Few-Shot Object Detection (AAAI 2024 · similarity 0.83)
- DiMSOD: A Diffusion-Based Framework for Multi-Modal Salient Object Detection (AAAI 2025 · similarity 0.83)
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.