YimianDai

CVPRW 2018, Oral @ MBCC Workshop

最后一个作者 John K. Tsotsos 的个人主页值得探索一下。

Priming

Priming, 中文叫作 "启动", 是一个蛮复杂的概念. 在这篇论文中, 作者认为 visual priming 就是 an effect of top-down signaling in the visual system triggered by the said cue, 我们就按照本文的定义来理解, 不去管什么内隐记忆.

首先, cue 是与 label 不同的存在, cue 是在训练 label 之外额外的信息. label 只有在 training 阶段才有, 在 test 阶段则是没有的; 而 cue 则是在所有阶段都是有的. 事实上, cue 其实是一类 top-down signal, 是来源于 external (exogenous) stimuli 的 top-down signal. 除了 cue 之外, top-down signal 也可以来自于 internal (endogenous) processes of reasoning.

YimianDai

Step 1: 在本地机上安装 Docker

在 Ubuntu 上安装 https://docs.docker.com/install/linux/docker-ce/ubuntu/ 安装 Docker

Step 2: 安装 nvidia-docker-plugin 为了能够在 docker 中使用 GPU

根据 https://github.com/NVIDIA/nvidia-docker/wiki 安装

Step 3: 从 docker hub 上 pull 相应的 mxnet 镜像

YimianDai

本文的 Motivation 是 a learnable nonlinear function with spatial connections 来 making the nonlinear activations more effective. 事实上, xUnit, a layer with spatial and learnable connections 也可以理解成跟 SENet, GENet 一样的 Attention 模块. 从下图看, xUnit 其实也就是跟 GENet 一样的模块, 这点在 GENet 的论文里也提到了.

对我而言, 本文最大的贡献是指出了 Nonlinear Activation 函数可以写成 Element-wise Multiplication 的形式

原始的 Nonlinear Activation 形式

YimianDai

NIPS 2018，作者就是 SENet 的作者。

CV 算法里面通常的模式

augment functions that perform local decisions with functions that operate on a larger context, providing a cue for resolving local ambiguities

using simple aggregations of low level features can be effective at encoding contextual information for visual tasks,

为什么用 context 往往能够提升 object detection 算法性能?

YimianDai

Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks

Enhancement of SSD by concatenating feature maps for object detection

DSSD: Deconvolutional single shot detector

xUnit: Learning a Spatial Activation Function for Efficient Image Restoration

Priming Neural Networks

YimianDai

augment SSD+Residual-101 with deconvolution layers to introduce additional large-scale context in object detection and improve accuracy, especially for small objects

improve detection accuracy 的方式

1. exploiting multiple layers within a ConvNet
   • 方式 1: combine feature maps from different layers of a ConvNet and use the combined feature map to do prediction

YimianDai

Enhancement of SSD by concatenating feature maps for object detection

SSD 的缺点:

1. each layer in the feature pyramid is used independently 由此导致了 the same object can be detected in multiple scales
   • 具体说明: a certain position of a feature map in a lower layer (say, Conv4-3) is activated. This information can affect entire scales up to the the last layer (Conv11-2), which means that the relevant positions in the higher layers have a good chance to be also activated 但是 SSD 目前没有这种约束
2. small objects are not detected well

相应的该怎么改善 SSD

YimianDai

ECCV 2016 - SSD: Single shot multibox detector

2017 - DSSD: Deconvolutional single shot detector

BMVC 2017 - Enhancement of SSD by concatenating feature maps for object detection

SSD 的不足

SSD 的改进

YimianDai

small-size object 性能不好表现在会有很多 small-size false positives, 背后的根源在于 the inadequacy of low-level features

the inadequacy of low-level features 具体又表现为: small receptive field sizes 和 weak semantic capabilities

这篇论文的贡献在于 demonstrates independent predictions from different feature layers on the same region is beneficial for reducing false positives.

We propose a novel light-weight previewer block, which previews the objectness probability for the potential regression region of each prior box, using the stronger features with larger receptive fields and more contextual information for better predictions.

The lack of contextual information leads to unsatisfactory performance of multi-scale detectors on detecting small objects.

YimianDai

查看 Azure for Student 的余额的网站: <www.microsoftazuresponsorships.com>

美国西部 没有 GPU, 美国西部 2 (West US 2) 是有 GPU 的

N 系列针对特定工作负荷提供三种特定产品 / 服务：

NC 系列专用于高性能计算和机器学习工作负荷。全新版本 NCv3 配备 NVIDIA Tesla V100 GPU。 ND 系列专用于进行深度学习的培训和推理方案。该系列使用 NVIDIA Tesla P40 GPU。全新版本 NDv2 配备 NVIDIA Tesla V100 GPU。 NV 系列采用 NVIDIA Tesla M60 GPU，支持功能强大的远程可视化工作负荷和其他图形密集型应用程序。