这是一个电视直播机器人项目 系统会跟踪主持人的脸部，使脸部位于屏幕中间 主要算法是Yolov3，在WIDER FACE上训练并在...

# YOLOv3-Based-Face-Detection-Tracking This is a robot project for television live. System will tracking the host's face, making the face in the middle of the screen. Main algorithm is Yolov3, trained on WIDER FACE and tested on FDDB benchmark. Extremely fine performance! ## ***SEE THE VIDEO***: https://v.youku.com/v_show/id_XMzc1OTk5MTg3Mg==.html?spm=a2hzp.8244740.0.0 # Thanks to ***https://github.com/experiencor/keras-yolo3*** I used part of this project, changed some codes of it. # How to USE ## Environmnet Requirements: I creadted this project on my DL desktop, its characteristics are as following: * |name|details| * |system| linux16.04/windows10| * |platform| anaconda2 or 3| * |GPU| GTX1080TI| * |CUDA| 9.0| * |CUDNN| 5.0| * |tensoflow-gpu| 1.8.3| Some other dependecies you need for DL may occur some version conflicts, just Google it! ## Here comes the code First, download my tensorflow graph file from here: https://pan.baidu.com/s/1FRxVacEraMQkIQnDYxnO-w Change the .pb file path called in the file ***loadtfpb.py***, then you can run it and see the magic! Another code ***load_tf_pb-control-robot.py*** is the file I used control my robot. You can isgore that. # Results ## Test with pics from the internet and myself :)    ## Compare with other detector: I also found a MTCNN detector. In comparison, YOLOv3 is more robust. In the following, the first image is MTCNN test result and the second one is YOLOv3 test.   Trained the models with 10 epoches on Wider-Face becnhmark and test them with FDDB benchmark. Here is the ROC curve. You can judge by yourself!  ***Have Fun!***