ICCV 2019 - This week computer vision experts from around the world are gathering in Seoul, South Korea for ICCV (International Conference on Computer Vision). Our staff, for example came in from our headquarters in Austin, TX and our office in Kuala Lumpur, Malaysia.
ICCV is a hotbed of innovative ideas and technology aimed at giving the gift of sight to machines to solve complex problems. Here is a great list of highlights from the papers presented.
A Cornucopia of Applications
Participants arrived with a host of innovative use cases all of which demand training data. One team developing facial recognition software is interested in acquiring data with 100 facial keypoints per face. We talked to two companies developing lumber grading algorithms to detect defects on various wood surfaces. This requires both specific domain knowledge and hyperspectral imaging capabilities on the tooling side.
Other exciting projects include semi supervised robotics, body size estimation for clothing fitting, semi supervised robotics, medical imaging, and instance segmentation model for identifying gemstones in a setting.
What do they all have in common?
Each project demands high quality training data. Did you know that 96% of data scientists report experiencing challenges with data quality and labeling? Even the most sophisticated model can be easily compromised if trained on data that is poorly labeled or does not accurately reflect the target values. As ML initiatives become more complex and use-cases more diverse, use case specific training data becomes more central to model development.
Video has become increasingly popular for training computer vision algorithms. Even videos of moderate length and resolution can generate large payloads once they are annotated with positions and classifications.
Consider this: the standard frame rate for film is 24 frames per second, 1440 frames per minute, and 86,400 per hour. That’s a lot of annotating! Just think of the number of hours required for an autonomous vehicle to learn how to navigate the complexity of a city street.
To meet both quality and scale requirements, annotators need advanced tools. Interpolation and object tracking, for example, allow workers to quickly bridge several frames of motion together and the ability to work simultaneously on longer videos. Facilitating concurrent work requires a enterprise-grade platform with caching capabilities to manage streaming with the annotation interface.
Annotated videos hold the future for training computer vision models. Check out our latest piece on How To Tackle Video Annotation.