FDDF: Frequency Decomposition and Spatial-Frequency Dual-Domain Fusion Network for Multi-Spectral Pedestrian Detection
This repository contains a reference implementation of the core modules of FDDF (Frequency Decomposition and Spatial–Frequency Dual-Domain Fusion Network) for multispectral pedestrian detection. The method is described in our paper:
X. Liu, G. Xie, X. Xie, and X. Xu,
"FDDF: Frequency Decomposition and Spatial-Frequency Dual-Domain Fusion Network for Multi-Spectral Pedestrian Detection",
The repository currently includes the four key building blocks of our dual domain fusion paradigm, as well as the content of data labeling, training, and data processing:
- fdfd: Frequency-Domain Feature Decomposition Module
- fsc: Frequency–Spatial Domain Feature Global Co-occurrence Module
- sdci: Spatial-Domain Cooperative Integration Module
- fsa:Frequency Spectrum Attention Operation Module
Baseline Code and External References
Our implementation is built on top of existing open-source multispectral pedestrian detection codebases. In particular:
The training and detection pipeline (VGG-16 backbone + SSD detector, data loading, augmentation, and loss functions) is adapted from the official implementation of MLPD (“Multi-Label Pedestrian Detector in Multispectral Domain”) [Kim et al., RA-L 2021].https://github.com/sejong-rcv/MLPD-Multi-Label-Pedestrian-Detection.git
Some utility functions (e.g., anchor generation, evaluation scripts) follow the design of standard SSD implementations in PyTorch.
You can quickly evaluate the results by running evaluation_stcript.py FDDF_desult. txt is the result of this article on Kaist KAIST-annotation.json is a validation set label KASIT_SENCHMARK.jpg is a comparison chart with other methods