https://blog.csdn.net/chenmaolin88/article/details/79357263
https://www.pianshen.com/article/1158403131/
https://www.cnblogs.com/zongfa/p/9663649.html
https://www.cnblogs.com/cx-99/p/12358933.html
*********************************************************************
https://github.com/gakkifan2020/models/tree/master/research/object_detection
https://github.com/gakkifan2020/models/blob/master/research/object_detection/g3doc/tf1_detection_zoo.md
https://github.com/gakkifan2020/models/blob/master/research/object_detection/g3doc/tf1.md
_________________________________________________________________________
labelimg 比 labelme 好
1 json_xml_transfer 文件夹是用来进行数据格式转换的
2 training 文件夹里的 *.config 是从 samples -> configs 文件夹里拷贝出来,用于配置训练参数的, 和所选的模型是一一对应的
3 如果有 nan 报错,删除 legacy training 中保存的模型, config 文件中 增大 batch_size , 调小 学习率,重新训练
使用 tensorboard 在浏览器打开查看训练进度:
进入object_detection/legacy 文件夹下:
tensorboard --logdir=training , 打开浏览器链接
4 生成需要的模型文件: object_detection/export_inference_graph.py
input_type image_tensor
pipeline_config_path /home/zhangwei/anaconda3/envs/tf114/lib/python3.6/site-packages/tensorflow/models/research/object_detection/training/ssd_mobilenet_v1_coco.config
trained_checkpoint_prefix /home/zhangwei/anaconda3/envs/tf114/lib/python3.6/site-packages/tensorflow/models/research/object_detection/legacy/training/model.ckpt-20000
output_directory /home/zhangwei/anaconda3/envs/tf114/lib/python3.6/site-packages/tensorflow/models/research/object_detection/zhangwei_detection
5 测试模型: object_detection/object_detection_tutorial.py
model name, path to labels , num classes 更改成自己的,download model部分都删去
object_detection_tutorial.py 是用于测试模型效果的, 读取的模型来自于 zhangwei_detection 文件夹
测试时如果不理想,可以修改 iou 阈值
min_score_thresh=0.8