- Face Attention Network: An Effective Face Detector for the Occluded Faces, which is submitted to ECCV 2018 as a co-first writer, that achieves state-of-art (hard AP: 88.5) on WIDER Face dataset. PDF: https://arxiv.org/pdf/1711.07246.pdf- SFace: An Efficient Network for Face Detection in Large Scale Variations, which is submitted to ECCV 2018 as a co-first writer, that efficiently integrates the anchor-based method and anchor-free method to address the scale issues in face detection. PDF: https://arxiv.org/pdf/1804.06559.pdf- First place in WIDER Face and Pedestrian Challenge 2018 (Workshop in conjunction with ECCV 2018) Track 1: Face Detection, that achieves mAP: 55.82, 2.5AP more than the second place.- Designed minimal FLOPS CNN structures, and assisted full-time employees to migrate the company's original face detection architecture to RetinaNet, which comprehensively improved the face detection performance of the mobile phone product line.- Designed the 4-direction rotating face detection model to raise the detection performance of the mobile phone face detection model in the 4-direction rotating face to a level close to the positive direction face, and assisted full-time employees to win several orders of the facial unlocking modules for mobile phone.

- Developed an ID card image quality control SDK on mobile phone, which can automatically estimate the quality of uploaded image, and increase the single-frame automatic audit pass rate to more than 95%. - Designed and implemented an ID card cloud error code interface, which can automatically identify the reason why the text recognition program failed on ID card.- Designed word level fusion strategy based on Baidu’s own OCR model, raised the F-score by 2.07% on ICDAR 2015 - Task 4.1.

- Core Member of the initial entrepreneurial team, responsible for the development of job recommendation system, including data collection, cleaning, feature extraction, recommendation, feedback and other processes.- Designed an end-to-end recommendation platform with online and offline modules.