方案简介
本方案基于京东FastReID框架实现。 网络采用ResNest101, 损失函数为circle-loss + triplet-loss.
系统安装
$ conda install pytorch==1.6.0 torchvision tensorboard -c pytorch
$ git clone https://github.com/wangbingo/fast-reid.git
$ cd fast-reid
$ pip install -r requirements
$ cd fast-reid/fastreid/evaluation/rank_cylib && make all
数据准备
$ cp XXXX/train.zip ./ && cp XXXX/image_B.zip ./
$ unzip train.zip && unzip image_B.zip
$ cd fast-reid && python tools/prepare_train-data.py && python tools/prepare_test-data.py
###将数据组织成market1501形式，便于直接利用框架
$ mkdir -p fast-reid/datasets/Market-1501-v15.09.15 && \
mv fast-reid/datasets/pclreid/train fast-reid/datasets/Market-1501-v15.09.15 && \
mv fast-reid/datasets/Market-1501-v15.09.15/train fast-reid/datasets/Market-1501-v15.09.15/bounding_box_train
$ mv /content/fast-reid/datasets/pclreid/* /content/fast-reid/datasets/Market-1501-v15.09.15
$ mv /content/fast-reid/datasets/Market-1501-v15.09.15/gallery /content/fast-reid/datasets/Market-1501-v15.09.15/bounding_box_test
训练
$ cd fast-reid && python ./tools/train_net.py --config-file ./configs/Market1501/sbs_S101.yml MODEL.DEVICE "cuda:0"
###训练完成后，权重文件保存在fast-reid/logs/market1501/sbs_S101
由于权重文件较大，故存放在百度盘供提取。
链接: https://pan.baidu.com/s/10gM1TKk2LuNUw086DKCWUg 密码: 1vr6
测试（test）
$ cd fast-reid/ && sh demo/my_run_demo.sh
###测试完成后，结果文件保存在fast-reid/result_2020-10-14-08-56-35.json(时间为当前系统时间)