Participated Research Projects

Visual Traffic Scnene Understanding for Autonomous Driving (CMU/GM)

This project is under the collaborative research between CMU and GM to develop advanced technologies for autonomous driving as well as future model year vehicles. At the core of this project is the goal of collision-free drivable space estimation, and the development of scene parsing and understanding algorithms centered on this theme.

CASENet: Deep Category-Aware Semantic Edge Detection (MERL)

This work addresses category-aware semantic edge detection using CNNs. Unlike binary edge detection, the problem requires one to jointly detect semantic edges and assign them with object category labels. We propose an end-to-end learning framework which achieves state of the art results on SBD and Cityscapes. The work was done during intern at MERL, with one paper submitted to CVPR 2017.

Deep CNN based Static Facial Expression Recognition (MSR)

This work was an intern project at MSR in 2015. The targets of this project were:
1) Develop a system able to robustly recognize facial expressions using CNN.
2) Participate the Third Emotion Recognition in the Wild Challenge (EmotiW 2015).
Our submitted results won 2nd place on the SFEW sub-challenge in EmotiW15, with one long paper accepted to ICMI15 as oral. A major portion of the methods and codes are integrated to the Emotion API under Microsoft Cognitive Services.

Image Selection and Voice Controlled Image Editing for PixelTone (Adobe)

During the 2013 summer intern at Adobe, I developed several new features for PixelTone - A voice controlled image editing prototype:
1) An interactive graph cut and quick matting tool for object/region selection.
2) An intelligent content-aware scribbling tool able to tolerate scribbling errors.
3) A natural language interaction system for color identification and indication. (e.g.: “What color is it? Make it a MILD SKY BLUE and INTENSE PEACH color…”)

Segmentation and Object Proposal for Large Scale 3D Urban Point Clouds (SIAT)

In this project, we conisder the problem of generating object proposals for 3D urban point clouds. The target is to propose a point cloud segmentation method that can automatically segment common objects from urban scenes, so that subsequent tasks such as object recognition and 3D urban modeling can benefit. This project is a joint work with Shenzhen Insititutes of Advanced Technology, Chinese Academy of Sciences.