San Francisco-based stealth startup PerceptiveIO publishes an open-access paper "UltraStereo: Efficient Learning-based Matching for Active Stereo Systems" by Sean Ryan Fanello, Julien Valentin, Christoph Rhemann, Adarsh Kowdle, Vladimir Tankovich, Philip Davidson, and Shahram Izadi.
"Mainstream techniques usually take a matching window around a given pixel in the left (or right) image and given epipolar constraints find the most appropriate matching patch in the other image. This requires a great deal of computation to estimate depth for every pixel.
In this paper, we solve this fundamental problem of stereo matching under active illumination using a new learning-based algorithmic framework called UltraStereo. Our core contribution is an unsupervised machine learning algorithm which makes the expensive matching cost computation amenable to O(1) complexity. We show how we can learn a compact and efficient representation that can generalize to different sensors and which does not suffer from interferences when multiple active illuminators are present in the scene. Finally, we show how to cast the proposed algorithm in a PatchMatch Stereo-like framework for propagating matches efficiently across pixels."
ETH Zurich publishes PhD Thesis "Deep Neural Networks and Hardware Systems for Event-driven Data" by Daniel Neil.
"This thesis introduces hardware implementations and algorithms that use inspiration from deep learning and the advantages of event-based sensors to add intelligence to platforms to achieve a new generation of lower-power, faster-response, and more accurate systems."
Qualcomm announces Snapdragon Neural Processing Engine (NPE) SDK running on Kryo CPU, Adreno GPU or Hexagon image processing DSP. Facebook announced plans to integrate the Snapdragon NPE into the camera of the Facebook app to accelerate Caffe2-powered AR features. By utilizing the Snapdragon NPE, Facebook can achieve 5x better performance on the Adreno GPU, compared to a generic CPU implementation.
"Mainstream techniques usually take a matching window around a given pixel in the left (or right) image and given epipolar constraints find the most appropriate matching patch in the other image. This requires a great deal of computation to estimate depth for every pixel.
In this paper, we solve this fundamental problem of stereo matching under active illumination using a new learning-based algorithmic framework called UltraStereo. Our core contribution is an unsupervised machine learning algorithm which makes the expensive matching cost computation amenable to O(1) complexity. We show how we can learn a compact and efficient representation that can generalize to different sensors and which does not suffer from interferences when multiple active illuminators are present in the scene. Finally, we show how to cast the proposed algorithm in a PatchMatch Stereo-like framework for propagating matches efficiently across pixels."
ETH Zurich publishes PhD Thesis "Deep Neural Networks and Hardware Systems for Event-driven Data" by Daniel Neil.
"This thesis introduces hardware implementations and algorithms that use inspiration from deep learning and the advantages of event-based sensors to add intelligence to platforms to achieve a new generation of lower-power, faster-response, and more accurate systems."
Qualcomm announces Snapdragon Neural Processing Engine (NPE) SDK running on Kryo CPU, Adreno GPU or Hexagon image processing DSP. Facebook announced plans to integrate the Snapdragon NPE into the camera of the Facebook app to accelerate Caffe2-powered AR features. By utilizing the Snapdragon NPE, Facebook can achieve 5x better performance on the Adreno GPU, compared to a generic CPU implementation.
AI News: Machine Learning for Stereo Depth Mapping, DNN Processor for Event Driven Sensors
Reviewed by MCH
on
July 28, 2017
Rating:
Reviewed by MCH
on
July 28, 2017
Rating:
No comments: