ThunderNet

face++最近发了一篇名为ThunderNet的paper,整篇paper不仅介绍了ThunderNet本身的结构、模块和表现,更细致地说明了背后的设计思想和原理。没有花哨的专业词汇和高深的数学理论,但却对detection任务本身有深入的思考和理解,可以说是一篇相当良心的paper。虽然还没有发布复现代码,反正我信了。

这篇paper介绍了一种轻量级的two-stage detector。从速度及表现两方面都超越了SSD这一类one-stage detector。总的architecture仍是标准的two-stage,由backbone提取特征,再由一个RPN输出proposals,根据proposals从feature maps里提取ROI,最后对这些ROI做classification和localization的refinement。但作者通过选用轻量化的backbone,高效的RPN和检测头(detection head)设计以及引入两种新的结构模块——CEM(Context Enhancement Module)和SAM(Spatial Attention Module),大幅度提高了two-stage detector的效率。下图为ThunderNet的整体architecture。

Depthwise Separable Convolution

This post explains a neural network operation called depthwise separable convolution, which trades off between latency and accuracy.

Depthwise separable convolution factorizes a standard convolution into a depthwise convolution and a pointwise convolution. Depthwise convolution captures spatial information of each feature map channel and pointwise convolution combines these information across all channels.

Gradient Harmonized Single-stage Detector

This post introduces the paper ‘Gradient Harmonized Single-stage Detector’, which tries to tackle hard example mining problem. In the end, we compare between GHM and Focal Loss.

This paper tries to handle the long-existing and well-known problems of one-stage detector: the imbalance between the number of positive and negative examples as well as that between easy and hard examples.

Conditional Log-Likelihood vs. Mean Squared Error

This post introduces conditional log-likelihood and mean squared error and shows the relationship between these two performance measures.

For many machine learning problems, especially for supervised learning, the goal is to build a system that can take a vector \bold{x}\in{\mathbb{R}^m} as input and predict the value of a scalar y\in{\mathbb{R}} as its output. To solve such problems, we can follow the steps:

  1. choose a model to predict a \hat{y} when given an input \bold{x}
  2. estimate the weights \theta = g(\bold{x}) that \hat{y} = f(\theta, \bold{x})
    1. define the performance measure (also known as loss function or cost function)
    2. design an algorithm that will improve the weights \theta in a way that reduces loss function

Pagination


© 2017. All rights reserved.

Powered by Hydejack v7.5.1