The Latest Trends in Computer Vision
Markable Exhibits @ CVPR 2017
From July 21 to July 26 we attended the Conference on Computer Vision and Pattern Recognition (CVPR), the premier computer vision conference in the world. The sheer size of the event is a testament to how much the industry has grown and how important it has become, not just to computing, but to business in general.
This year’s conference was held in Honolulu, and 5,000 researchers, engineers and business leaders attended — almost 50% more than a year ago and 4900% more than the first CVPR conference in 1983, which had only 100 attendees. This year, 2,620 valid papers were submitted with 783 accepted. This is up from 2,145 and 643, respectively, in 2016. And, to top it off, 30% more sponsors supported this year’s event than in 2016. Needless to say, this conference has become a big deal.
We were excited to catch up with old friends and meet some of the pioneers in the field. We met with investors, hardware manufacturers and fashion retailers, all of whom showed interest in working with Markable. And of course, many students approached us hoping for a chance to work with us on such an exciting new problem. Overall, this year’s conference was buzzing with the energy of an industry on the verge of explosive growth.
Past conferences focused on the potential of AI and where it was headed. But the future is now, and we are starting to see the technology’s promise manifest in practical applications that are disrupting industries from intelligent video analytics and self-driving cars to medical imaging and advertising.
The main topics covered this year were Machine Learning, Object Recognition & Scene Understanding, 3D Vision, Human Analyzing, Neural Networks and Video Analysis; all of which are important to the foundation of Markable’s algorithms.
Better design of neural networks helps to improve the performance on not just the various computer vision problems, but also the other broader machine learning problems. Related to this, we liked the DenseNet paper from Facebook AI Research group which improves on the ResNet network to enable learning in deeper networks via shortcut connections.
In terms of research along object detection, we liked Feature Pyramid Network which proposes a better way to model top-down and bottom-up context to improve detection for smaller objects. It was encouraging to see more research closer to home in fashion recognition. One of the noticeable papers was “Video2Shop: Exact Matching Clothes in Videos to Online Shopping Images” by Alibaba Group. We await the public release of their dataset so that we can benchmark our progress against this paper.
Object recognition was an especially hot topic this year and an area in which we excel. Past work was focused on recognizing single objects and understanding the scene as a whole. Now, the focus has shifted to understanding the relationship between multiple objects in a single image. This is something Markable has been working on for some time and an area where we are unmatched in relation to fashion.
Another critical issue discussed at length at this year’s event was data. Data has become the logjam of advanced algorithms. Deep learning requires massive amounts of data and it is imperative for our success at Markable. The quality of data is also critical, since low-quality data severely hinders performance. It won’t matter how good the model is, if the data is flawed, the outcomes will be too.
After five days of meeting so many brilliant people and listening to the results of inspired research, it was encouraging to see that Markable remains at the forefront of computer vision. It is an industry that is only in its infancy. But judging by the attention it is receiving from global tech titans and researchers alike, we know it will grow up fast. And we are excited for what is to come.