In June 2019, researchers from Ciklum’s R&D department visited the Computer Vision and Pattern Recognition (CVPR) Conference held in Long Beach, California. CVPR is an annual event on computer vision and pattern recognition, which has been held every year since 1983. CVPR is one of the world’s top three academic conferences in the field of computer vision. This year Ciklum research engineers visited the conference not only as attendees but as presenters. Here are some highlights from the sessions.
Last year the conference attracted more than 9000 attendees. The number of submitted papers is no less impressive at 5160.
The number of papers submitted and accepted, 2015-2019
It’s also worth mentioning that CVPR 2019 received a record high of over $3.1 million in sponsorships.
Ciklum`s senior research engineer, Oleg Panichev, gave a presentation titled “U-Net Based Convolutional Neural Network for Skeleton Extraction” at CVPR’s “Deep Learning for Geometric Shape Understanding” workshop. In his presentation, Oleg shared how he together with a team of researchers developed a deep learning model that could extract the skeleton of different objects with high precision. The model developed by the team enabled them to win the Pixel SkelNetOn Competition. Here are some examples of objects from Oleg’s slides:
Oleg Panichev is giving a presentation at CVPR 2019
HIGHLIGHTS FROM CVPR 2019
Here are the most insightful workshops and presentation topics according to the researchers from Ciklum’s R&D team:
Person re-identification is a basic task for intelligent video surveillance systems. Simply stated, it involves tracking the same person across a network of images/videos collected from multiple cameras and at different times. Given the large volume of surveillance data, finding a method that can efficiently extract meaningful information is a challenging task. Read about solutions proposed during CVPR here.
Semantic segmentation assigns pixels to classes, but it does not involve the identification of individual instances inside an image as with instance segmentation. Thus, the task of panoptic segmentation is basically to combine them to work simultaneously. Learn about the metrics proposed for more unified image segmentation here.
Listen to the Image
Cross-modal plasticity theory claims that the loss of one sensory modality leads to enhanced sensory performance in remaining ones. In other words, it is possible to use other senses (e.g., hearing) as the sensor to “visually” perceive the environment. The devices used to compensate for lost senses through other sensory channels is called Sensory Substitution (SS) devices, which are life-changing for the blind and deaf.
The device presented at the conference got name, “The vOICE”. The camera on the forehead of a blind person captures the scene in front, which is transformed into a sound that conveys the visual information. Check the results here.
Machine learning systems are poised to grow and transform how we work and interact. CVPR is a great place to get first-hand insights about the future from top machine learning experts. The papers presented at the event help them further develop visual recognition, object tracking and object labelling skills.
Modern problems require state-of-the-art solutions, so if your business needs guidance on how to implement machine learning, contact Ciklum.