Multi-CNN feature ensemble method.
- We have trained 4 GoogLeNet with different task, e.g, triplet loss (full image), vehicle_id softmax loss (full image), vehicle_id softmax loss (upper half image) and model_id softmax loss (lower half image).
- In inference stage, we concat the
pool5/7x7_s1
layer feature from the 4 GoogLeNet together. Finally, return thek
nearest vehicles.
We used the method in Vehicle Retrieval task of The 3rd National Gradute Contest on Smart-CIty Technology and Creative Design, China. We ranked 1st and won the special prize in the final!
The Dataset used in Vehicle Retrieval task: PKU VehicleID. Note: if you want to use the dataset, go to the website and ask for the download link.
- CPU or GPU: CPU only
- OS: Windows x64
- DL tool: Caffe
- Compiler: VS2013
- Windows Caffe (use this version) setup.
- Download or git clone the current project.
- Copy or move
vs_vehicle_retrieval_kCNNs
folder intocaffe/windows
and addvehicle_retrieval_kCNNs.vcxproj
project into Caffe solution in VS2013, compile it with Release mode. - Modify
run.bat
, mainly set the path. Finally, runrun.bat
in cmd, you'll get axml
result file.
第三届全国研究生智慧城市技术与创意设计大赛车辆精确检索任务第一名,总决赛特等奖。
数据集:PKU VehicleID
基于深度学习的多模型集成方法。
CPU,Windows系统,Caffe,VS2013
- 下载、配置、编译Caffe官方windows版(https://github.com/BVLC/caffe/tree/714d0acad8c66d64ddf7b83b9a239f7efc017894)
- 下载本工程
- 将文件夹
vs_vehicle_retrieval_kCNNs
复制到caffe/windows
目录下,并在vs中把vehicle_retrieval_kCNNs.vcxproj
项目添加到Caffe解决方案下,使用Release模式编译生成可执行文件。 - 修改
run.bat
中的路径,运行它,即可得到实验结果。