Face Antispoofing
Face liveness detection alias antispoofing is an important application of deep learning for online registration, especially to banks, stock company etc. It mainly consists of face detection and facial landmark detection module, followed by face recognition. This is an android implementation of face-antispoofing function.
Ultimate Face Detector
Previous face detectors are not as robust as we want them to be in terms of large pose-variations, occlusions, illumination changes and small faces. Therefore, I've trained a powerful a face detector that can handle the above issues with near real time efficiency. The detector achieves over 95% mAP on FDDB and is yet to be incorporated in a commercial application framework.
Reinforcement Learning
I only knew how to use A* and IDA* searching algorighms for solving the maze problem when I was participating ACM proramming competition. Now, thanks to the help of Reinforcement Learning, I could devise deep Q network to choose action based on the feedback of the environment.
Scrapy the imdb
To get sufficient data for CNN training, I have made an attempt to scrap the imdb using Scrapy. I've never used Regular Expression so frequently before! The reference image of person is first acquired and then his/her associated pics. One tricky part is to use Selenium to simulate JS for generating requests.
Two Stage GAN
Although much progress has been made in Generative Adversarial Network, this one is considered to be another milestone in its latest development. It achieves the adverse function of Image Captioning--ie. given a text, output an image according to the description. The implementation of two-stage GAN involves the combination of Torch and Tensorflow, which is really exhilarating.
Video Recognition
The academic and industrial community has seen a wide range of applications of CNN in image tasks, here we try CNN on the task of video recognition. Here we use the two-stream Convolultional network to capture the spatial stream and temporal stream respectively.