NoScope:Optimizing Neural Network Queries over Video at Scale
应用场景:二分类查询,即判断某一帧有没有某一对象。
解决的问题:对象查询处理
描述:NoScope针对神经网络查询昂贵的问题,提出了使用专用NN和跳帧的方法,以及使用优化器在速度和精确度之间进行平衡,保证在用户所需精确度的前提下,花费最少。
Motivation:
(1)Applying general NNs to video data in real time is computationally expensive at scale.
eg:80 fps on an NVIDIA P100, ~$4600 retail
(2)General NNs continue to get deeper and more costly to evaluate(higher computational cost)
Goal:
apply NNs to video analytics at a substantially lower computational cost,主要从下面两方面进行考虑:
(1)The complexity of the model
(2)The data size of calculation
Problems:
(1)how to reduce the complexity of the model?
(2)how to reduce the data size of calculation?
Solution:
(1)Model Specialization
(2)Difference Detector
(3)Cost-Based Optimizer
系统开源