这个项目是对原tensorflow models的一个剥离版本，原来的感觉有点乱，在这里我会一直维护，基本上会保持和官方repo一样的更新。

官方提供了SSD， Faster-RCNN的预训练模型，并且使用了不同的特征提取网络。从这里可以找到对应的下载链接。目前支持这些组合：

• SSD_MobileNet
• SSD_Inception_V2
• RFCN_ResNet101
• faster_rcnn_ResNet101
• faster_rcnn_Inception_ResNet_v2

可想而知，最后一个模型最大也最慢，但是精度是最高的，第一个速度最快，下面的预测结果就是SSD和MobileNet预测的：

我整理了一个一键预测的代码，从 test_images/ 连续预测两张图像，可以稍微修改一下就可以直接预测单张图片或者调用camera进行预测。速度感觉…还可以，我的是CPU。预测代码是 object_detection.py.

有时间我在把训练KITTI数据集的部分写完，其实这很简单，就是把图片转成tfrecord然后写proto修改参数。不过说实话，检测网络普遍比较复杂。

To be honest, Google's protobuf is a great thing, but it's not something just like json or xml. In this project, a lot of protobuf messages has been used. Many data models were converted into protos, so how to generate them is very important.

You gonna download protobuf git repo and build it from source, after that, you will probably got protoc the compiler. Just inside the ./tensorflow_models_detection

protoc -I ./ --python_out ./ ./protos/anchor_generator.proto

# actually you can convert all protos into python file
protoc -I ./ --python_out ./ ./protos/*.proto

the option -I is also called --proto_path, this indicates the proto save path as well as the proto search path (sometimes one proto may import from another proto, so how to find them you should specific this path, by usually place them into one directory like protos is more make sense.).

本文由在当地较为英俊的男子原创，转载请注明出处。如有疑问或者想交流，可以加入微信: jintianiloveu