Pose Estimation, Queues, and Input Pipelines in TensorFlow

by Sophia TurolJune 8, 2016
Learn about the challenges of pose estimation—poor image quality, occluded pose elements, etc.—and how to overcome them using fully convolutional networks.


Below are the videos from the TensorFlow San Francisco meetup—sponsored and organized by Altoros on May 17, 2016.


Pose estimation with TensorFlow

In his session, Alex Londeree of Knit Health talks about the challenges of pose estimation and the possible ways out. For instance, he mentions such difficulties as the poor quality of the images, the necessity to work on the occluded pose elements, etc. To overcome these obstacles, the Knit Health’s team pre-trains a model with an FCN architecture, making use of specialized spatial regressors.



TensorFlow queues and input pipelines

Yaroslav Bulatov, previously working for Google and now as an independent ML consultant, talks about parallel execution (with Python, inter alios), TensorFlow queues / queue runners, and everything in between.



Fireside chat

At the wrap-up of the meetup, Alex and Yaroslav answered some questions from the audience, regarding fields of TensorFlow use, TensorBoard, etc.



Join our group to stay informed about the upcoming meetups!


Further reading


About the speakers

Alex Londeree is Senior Data Scientist at Knit Health. He is primarily training deep learning networks to do the necessary computer vision tasks for the company’s products. Alex has a background in deep learning and distributed machine learning systems. Using TensorFlow, he managed to design, implement, and deploy an embeddable deep convolutional neural network for pose estimation and object recognition in a real-time video.


Yaroslav Bulatov is an independent machine learning consultant with experience in implementing and training large-scale neural networks for Google Brain. He designed the first system that outperformed humans at recognizing outdoor house numbers. Yaroslav has a degree in computer science from Oregon State University, where he researched hand recognition using geometric hand classifiers. Yaroslav is a coauthor of a paper on training conditional random fields—used in part-of-speech tagging, text-to-speech mapping, protein and DNA sequence analysis, as well as information extraction from web pages.