Apply for Trial
Newsroom

SenseTime Scores Three Wins in World-Class Computer Vision Challenges at CVPR 2020

2020-06-18

5eeb114d2d179.jpg


Hong Kong – June 18, 2020 – SenseTime, a leading global artificial intelligence (AI) company, is proud to announce winning three world-class challenges in computer vision at the Conference on Computer Vision and Pattern Recognition (CVPR), following its acceptance of 62 research papers to the conference. SenseTime was ranked first in two challenge tracks of the International Challenge on Activity Recognition (ActivityNet), and also came first in another challenge track of New Trends in Image Restoration and Enhancement (NTIRE) this year.

 

ActivityNet is one of the world’s most prestigious competitions on activity recognition which plays an important role in the development of video analysis and liveness detection applications. SenseTime won two challenge tracks under ActivityNet, namely the AVA (Spatio-temporal Action Localization) and Kinetics Challenge (Trimmed Activity Recognition).

 

The X-Lab of SenseTime Research topped the AVA track of the AVA-Kinetics Crossover Challenge. The track was intended to evaluate the ability of algorithms to localize human actions in space and time, using the AVA-Kinetics Dataset. SenseTime’s algorithm scored a mean Average Precision (mAP) of 39.62, which was far ahead of the closest runner-up who only scored 6.71mAP.

 

SenseTime and its joint CUHK lab was also the joint winner at the Kinetics Challenge, a track under the ActivityNet, which evaluates the ability of algorithms to recognize activities in trimmed video sequences by using Kinetics dataset. SenseTime leveraged its deep-learning supercomputing platforms to train multiple video classification models with ultra-deep networks and was able to track and understand all the human actions in the video clips in a very tight timeline.

 

In the NTIRE challenge, SenseTime’s research team was crowned champions in the Spectral Reconstruction Challenge. SenseTime’s innovative level-4 network structure was able to expand the receptive field while extracting features from subnetworks at different levels to recover hyperspectral data from uncompressed 8-bit RGB images. The research provided a new solution to low-level vision problems for image restoration, enhancement and manipulation in the computer vision area.

 

Since its establishment in 2014, SenseTime has won more than 60 championships in different computer vision competitions, which is also a demonstration of its research capabilities and relentless pursuit of original and innovative computer vision technologies.

 

With these research achievements, SenseTime will continue its commitment to advancing AI technologies by creating new algorithms, developing open-source software and models, and advancing research and development in computer vision. In the meantime, SenseTime will also devote itself to transforming cutting-edge research ability into impactful business applications, empowering various industries and businesses for a better tomorrow.