Research

This page is under heavy construction. Only the most recent projects are listed here. Please go to the publications page for a full list of our papers and projects.




Realtime HEVC Encoder

 

With a flexible software defined architecture on general purpose computational platform(multicore CPUs and GPUs), we take a leading role to developed a realtime HEVC encoder - ZenHEVC, which currently support 1080p@60fps and 4K@30fps with high compression performance. Compared to HEVC reference code, HM10.0, the ZenHEVC is less than 0.5dB in terms of RD loss while maintaining realtime HD encoding speed. ( More...)


4K Video Sequences Quality Evaluation


A set of 15 rawdata sequences in formats 10 bits YUV 4:4:4 and 8 bits YUV 4:2:0 with frame rate 30 are availabe for researches incluiding HEVC, QoE and so forth. A full dataset about sujective scores for 4K@30fps has been released.( More..)


Video Stabilization and Performance Evaluation


The project focuses on improving the efficiency of video stabilization algorithms as well as developing better deshaking assessment tools to identify shaking and evaluate the exsiting algorithms. We have proposed a video stabilization algorithm based on L1-L2 optimization and built online demo, Android APP for cloud-based stabilization.(More...)


Deep Learning for Image and Video Processing


Deep learning(DP) has become a Swiss Army knife for machine learning and related research in the era of big data since its great success on ImageNet challenge in 2012. We are working on DP for modern image and video processing like video annotation(More...), Image Captioning(More...)...

SJTU-Spark


Spark is an open-source cluster computing framework originally developed in the AMPLab at UC Berkeley. Its in-memory primitives provide performance up to 100 times faster for certain applications. We have utilized spark in data mining, streaming engine for massive datasets online, and plan to implement machine learning algorithms on Spark.(More...)