Because there is no tutorial online about SSD training our own data, in this blog, I will explain how to train our own data. I hope this blog can solve the doubts about SSD.
Tips: This blog is my last blog (SSD's configuration installation and testing) of the sub blog, you need to install my last blog, which address is http://blog.csdn.net/samylee/article/details/51822832
Step one:Production of VOC2007 datasets
I made the pedestrian datasets, pictures to reach around18000,annotations to reach about100000. Due to the need of engineering company, the datasets is not convenient to display. I hope you will forgive me.
The process of making the datasets refer to this blog: http://blog.csdn.net/samylee/article/details/51201744, please.
The results of the production are stored in the data folder, named VOC2007, which contains three folders(Annotations, ImageSets, andJPEGImages). The format is the same as the Faster-rcnn.
Step two: Modify files of create_list.sh andcreate_data.sh
The download link of modified file of create_list.sh is: http://download.csdn.net/detail/samylee/9572820
The download link of modified file of create_data.sh is: http://download.csdn.net/detail/samylee/9572823
Step three: Modify files of ssd_pascal.py and ssd_pascal_webcam.py
The download link of modified file of ssd_pascal.pyis: http://download.csdn.net/detail/samylee/9572825
The download link of modified file of ssd_pascal_webcam.py is: http://download.csdn.net/detail/samylee/9572834
Step four: Experimental description
The experimental procedure is the same as my last blog(SSD's configuration installation and testing), we just need to modify the corresponding file to train our own data.
Step five: Experimental result
Iteration 60000, loss = 1.5299 Detection_eval = 0.7737
任何问题请加唯一QQ2258205918(名称samylee) !!!
或唯一VX:samylee_csdn