Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
our
history
meet
kuri
pet & person detection
place recognition
moment selection
face detection
pet &
person
detection
Kuri’s embedded GPU: if we want to process 5 frames a second, we can
only use ~2.5 GMACs/frame
MobileNets: Efficient Convo...
place recognition
•Uses feature vectors from intermediate layer of object detection network
•Cosine distance of vectors -> difference score
...
moment selection
X
X
face detection
Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks
Zhang, Kaipeng. et al. ...
BEEP BLOOP
KAIJEN HSIAO | ROBOTICIST & CTO | @KURIROBOT
Kaijen Hsiao at AI Frontiers: Adorable Intelligence
Kaijen Hsiao at AI Frontiers: Adorable Intelligence
Kaijen Hsiao at AI Frontiers: Adorable Intelligence
Kaijen Hsiao at AI Frontiers: Adorable Intelligence
Kaijen Hsiao at AI Frontiers: Adorable Intelligence
Kaijen Hsiao at AI Frontiers: Adorable Intelligence
Kaijen Hsiao at AI Frontiers: Adorable Intelligence
Upcoming SlideShare
Loading in …5
×

Kaijen Hsiao at AI Frontiers: Adorable Intelligence

406 views

Published on

Join us to hear how we created Kuri, the world’s most adorable home robot, and how cuteness and machine learning algorithms come together in creating a robot that people are excited to bring into their homes. Kuri uses embedded, deep-learning-based algorithms for face, pet, and person detection, as well as for place recognition (for mapping and localization). Such algorithms are crucial for enabling her endearing interactions with people, as well as her ability to be the family videographer and entertainer. Cute behaviors also enable Kuri to subtly and smoothly gather necessary data for training and inference, providing a significant benefit in improving core functionality for adorable home robots.

Published in: Data & Analytics
  • Be the first to comment

  • Be the first to like this

Kaijen Hsiao at AI Frontiers: Adorable Intelligence

  1. 1. our history
  2. 2. meet kuri
  3. 3. pet & person detection place recognition moment selection face detection
  4. 4. pet & person detection
  5. 5. Kuri’s embedded GPU: if we want to process 5 frames a second, we can only use ~2.5 GMACs/frame MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications Howard, Andrew G. et al. arXiv preprint 2017 (all authors from Google) Framework Model mAP Billion Mult-Adds Single Shot Multibox Detector 300x300 VGG 21.1% 34.9 Inception V2 22.0% 3.8 MobileNet 19.3% 1.2 pet & person detection
  6. 6. place recognition
  7. 7. •Uses feature vectors from intermediate layer of object detection network •Cosine distance of vectors -> difference score Useful for: Mapping loop closures Global localization “ConvNet features for Place Recognition,” Sunderhauf et al, IROS 2015 place recognition
  8. 8. moment selection X X
  9. 9. face detection Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks Zhang, Kaipeng. et al. Signal Processing Letters, 2016
  10. 10. BEEP BLOOP KAIJEN HSIAO | ROBOTICIST & CTO | @KURIROBOT

×