This project is for personal study and under development, please click 'watch' or 'star' my repo and check back later if you are interested in it.
主要任务:在交通场景下,实现对道路目标(车辆和行人)的实时检测。
目录结构:
FCOS
├── configs 配置
| ├── kitti_config.py
| └── bdd100k_config.py
├── data 数据
| ├── kitti.py
| ├── bdd100k.py
| ├── transform.py 数据变换
| └── collate.py 数据打包
├── models 模型
| ├── backbones 特征提取网络
| | ├── vgg.py
| | ├── resnet.py
| | ├── darknet.py
| | ├── mobilenet.py
| | ├── shufflenet.py
| | └── efficientnet.py
| ├── necks 特征融合网络
| | ├── fpn.py
| | ├── pan.py
| | └── bifpn.py
| ├── layers 网络模块
| | ├── conv.py
| | ├── spp.py
| | ├── aspp.py
| | ├── se.py
| | └── cbam.py
| ├── head.py 检测网络
| ├── target.py 训练目标
| ├── loss.py 损失函数
| ├── detect.py 检测后处理
| └── fcos.py 完整网络
└── tools 工具
├── train.py 训练
├── test.py 测试
├── eval.py 评估
├── infer.py 推理
└── demo.py 演示
本项目基于Anchor-Free的FCOS算法构建模型。
论文:https://arxiv.org/pdf/2006.09214.pdf
代码:https://github.com/tianzhi0549/FCOS
本项目基于公开的KITTI和BDD100K数据集训练模型。
官网:http://www.cvlibs.net/datasets/kitti/index.php
论文:http://www.cvlibs.net/publications/Geiger2013IJRR.pdf
目录结构:
kitti
├── training
| ├── image_2
| | ├── 000000.png
| | └── ...
| └── label_2
| ├── 000000.txt
| └── ...
└── testing
└── image_2
└── ...
统计信息:
-
类别数:8
-
类别名称:Car, Van, Truck, Pedestrian, Person_sitting, Cyclist, Tram, Misc
-
场景:City, Residential, Road, Campus, Person
-
训练集图片数:7481
-
测试集图片数:7518
-
图片分辨率:1224x370、1238x374、1242x375、1241x376
-
图片宽高比:3.3:1
官网:https://bdd-data.berkeley.edu/
论文:https://arxiv.org/pdf/1805.04687.pdf
目录结构:
bdd100k
├── images
| └── 100k
| ├── train
| | ├── 0000f77c-6257be58.jpg
| | └── ...
| ├── val
| | └── ...
| └── test
| └── ...
└── labels
└── 100k
├── train
| ├── 0000f77c-6257be58.json
| └── ...
└── val
└── ...
统计信息:
-
类别数:10
-
类别名称:Bus, Light, Sign, Person, Bike, Truck, Motor, Car, Train, Rider
-
时间:Dawn/Dusk, Daytime, Night
-
天气:Clear, Partly Cloudy, Overcast, Rainy, Snowy, Foggy
-
场景:Residential, Highway, City Street, Parking Lot, Gas Stations, Tunnel
-
训练集图片数:70k (137张缺少标注)
-
验证集图片数:10k
-
测试集图片数:20k
-
图片分辨率:1280x720
-
图片宽高比:1.78:1
本项目基于准确性和实时性指标评价算法性能。