[this repo is a Google Translate version of Windaway/DSOD-Tensorflow]
Using Tensorlayer and Skimage, etc., the modified Tensorlayer is used in the process. The pre-upload modification should be able to run in the original TL without testing the code. According to the paper, you can train on VOC07+12, and Los Loc gives high weight.
The source framework referenced the work of jasonli8848d and made some changes. Mainly Groundtruth that there is augmentation. In addition, the augmentation of this code consumes a lot of CPU resources, the training efficiency is very low, and the convergence is not fast. If you have any questions, please go to a comment or email me, colinyoo#outlook#com. Although it may not be able to help.