Talking Autonomy: Why Computer Vision Can't Solve the Long Tail скачать в хорошем качестве

Talking Autonomy: Why Computer Vision Can't Solve the Long Tail

Ghost Autonomy co-founder and CTO Volkmar Uhlig unpacks one of the biggest challenges in autonomy to date - the "long tail" of AI-based image recognition and object classification. Most approaches to autonomous vehicles involve some form of AI-based camera perception algorithms, and typically those algorithms use image recognition and object classification as a fundamental step. The approach is relatively well-understood: show a computer hundreds of thousands or millions of images of what might be on the road, and the neural network can be trained to recognize cars, buses, trucks, motorcycles, humans, etc. The problem: how do you train a network completely when there could be literally anything on the road? And how can you know that you've sufficiently trained all objects in every state of lighting, rotation, size, color, and partial reveal? Volkmar discusses these challenges and why Ghost believes a new approach to perception is required for autonomous driving. Thank you for watching Talking Autonomy, a series of short tech talks that explain the key elements of our work at Ghost. Subscribe and stay tuned for new episodes as we continue exploring the core technologies behind the Ghost Autonomy Engine and sharing insights from the founders, engineers, designers, mathematicians, and even the policy-makers who are responsible for bringing Ghost’s self-driving technology to the roads. Connect with Ghost and learn more: Ghost Autonomy Engine on the web: https://www.ghostautonomy.com/platform KineticFlow vision neural network: https://www.ghostautonomy.com/kinetic... Follow Ghost on Twitter:   / ghostautonomy   Follow Ghost on LinkedIn:   / ghostautonomy   #selfdriving #AI #autonomous #GhostAutonomy #computervision #objectrecognition