FAD is a object detection network.

• Overview
• System Requirements
• Installation Guide
• Dataset
• License

Overview

FAD aims to achieve efficient detection of traffic scene objects. Given the pressing demand for traffic object detection in autonomous driving vehicles, attaining efficiency and accuracy in perception on in-vehicle platforms remains a formidable challenge. To this end, we propose a frequency-oriented adaptive detector (FAD) for vehicle-mounted intelligent traffic object detection. FAD effectively integrates high- and low-frequency information to enhance the texture and semantic features of multi-scale objects in complex traffic scenarios, all while maintaining minimal computational overhead. Initially, we present a frequency dynamic convolution (FDC) for a lightweight backbone, aiming at frequency-domain adaptive dilated receptive fields and balancing effective bandwidth. Wherein, the adaptive kernel decomposes the convolution weights into high- and low-frequency components. Subsequent to frequency-domain selection, the convolutional weights are re-calibrated to equilibrate the frequency-domain components within the feature map. In addition, we propose an adaptive frequency-oriented fusion (AFF) framework, which adaptively reorganizes high and low-frequency features across different scales. AFF effectively balances object details, fine boundaries and deep semantic features, aiming to bridge the gap caused by feature inconsistencies.

System Requirements

Hardware requirements

FAD is conducted in the PyTorch framework on a Windows system, utilizing a workstation equipped with an RTX 4070 Ti 12-GB GPU and 32-GB RAM.

Software requirements

OS Requirements

This package is supported for Windows and Linux. The package has been tested on the following systems: + Windows 10 + Linux: Ubuntu 16.04

Python Dependencies

FAD mainly depends on the Python scientific stack.

``` basicsr==1.4.2 beautifulsoup4==4.13.4 einops==0.8.1 gradcam==1.5.4 ipython==8.12.3 matplotlib==3.5.3 mmcvfull==1.7.2 numpy==1.21.6 onnx==1.14.1 onnxruntime==1.19.2 opencvcontribpython==4.7.0.72 opencvpython==4.9.0.80 openvino==2023.2.0 pandas==1.3.5 Pillow==9.5.0 Pillow==11.3.0 protobuf==6.31.1 psutil==5.9.8 pycpuinfo==9.0.0 pycocotools==2.0.6 PyWavelets==1.4.1 PyYAML==6.0.1 PyYAML==6.0.2 scipy==1.7.3 seaborn==0.13.2 sentrysdk==1.39.1 setuptools==60.2.0 setuptools==65.6.3 thop==0.1.1.post2209072238 timm==1.0.17 torch==1.9.0+cu111 torchdct==0.1.6 torchsummary==1.5.1 torchvision==0.10.0+cu111 tqdm==4.65.2 torch==1.9.0+cu111

```

Installation Guide:

Install from Github

git clone https://github.com/feekyzz/FAD cd FAD conda create -n FAD python=3.8 anaconda conda activate FAD pip install -r requirements.txt - sudo, if required

Dataset

KITTI:

You can download KITTI from the following links:

• Google drive

Link: https://drive.google.com/drive/folders/1Ha6b9a1ri0VtZ4t5QCSMPaFHNWBScAC2?usp=sharing

BDD100K:

You can download KITTI from the following links:

• Google drive

Link: https://drive.google.com/drive/folders/1-gxQox4in4zzxcOgo66xN3BPN7fJbdIr?usp=sharing

Cityscapes:

You can download KITTI from the following links:

• Google drive

Link: https://drive.google.com/file/d/14wKLxFjgBVAdS4U_k0PKynrfJiLKbNdO/view?usp=sharing

Weights

• Google drive

Link: https://drive.google.com/drive/folders/1bgHoK5ASSymYoppbXnHdhND8rRYUTnpB?usp=sharing

License

This project is covered under the Apache 2.0 License.