Deep Learning for Computer Vision with Python

Description

Welcome to the ImageNet Bundle of Deep Learning for Computer Vision with Python, the final volume in the series. This volume is meant to be the most advanced in terms of content, covering techniques that will enable you to reproduce results of state-of-the-art publications, papers, and talks. To help keep this work organized, I've structured the ImageNet Bundle in two parts. In the first part, we'll explore the ImageNet dataset in detail and learn how to train state-of-the art deep networks including AlexNet, VGGNet, GoogLeNet, ResNet, and SqueezeNet from scratch, obtaining as similar accuracies as possible as their respective original works. In order to accomplish this goal, we’ll need to call on all of our skills from the Starter Bundle and Practitioner Bundle. The second part of this book focuses on case studies – real-world applications of applying deep learning and computer vision to solve a particular problem. We'll first start off by training a CNN from scratch to recognition emotions/facial expressions of people in real-time video streams. From there we’ll use transfer learning via feature extraction to automatically detect and correct image orientation. A second case study on transfer learning (this time via fine-tuning) will enable us to recognize over 164 vehicle makes and models in images. A model such as this one could enable you to create an “intelligent” highway billboard system that displays targeted information or advertising to the driver based on what type of vehicle they are driving. Our final case study will demonstrate how to train a CNN to correctly predict the age and gender of a person in a photo. 1 Introduction......Page 15 2 Introduction......Page 17 3.1 How Many GPUs Do I Need?......Page 19 3.2 Performance Gains Using Multiple GPUs......Page 20 3.3 Summary......Page 21 4.1.1 ILSVRC......Page 23 4.2.2 Downloading Images Programmatically......Page 25 4.2.4 ImageNet Development Kit......Page 26 4.2.5 ImageNet Copyright Concerns......Page 27 4.3 Summary......Page 29 5.1 Understanding the ImageNet File Structure......Page 31 5.1.1 ImageNet “test” Directory......Page 32 5.1.2 ImageNet “train” Directory......Page 33 5.1.3 ImageNet “val” Directory......Page 34 5.1.4 ImageNet “ImageSets” Directory......Page 35 5.1.5 ImageNet “DevKit” Directory......Page 36 5.2.1 Your First ImageNet Configuration File......Page 39 5.2.2 Our ImageNet Helper Utility......Page 44 5.2.3 Creating List and Mean Files......Page 48 5.2.4 Building the Compact Record Files......Page 52 5.3 Summary......Page 54 6 Training AlexNet on ImageNet......Page 55 6.1 Implementing AlexNet......Page 56 6.2 Training AlexNet......Page 60 6.2.1 What About Training Plots?......Page 61 6.2.2 Implementing the Training Script......Page 62 6.3 Evaluating AlexNet......Page 67 6.4 AlexNet Experiments......Page 69 6.4.1 AlexNet: Experiment #1......Page 70 6.4.2 AlexNet: Experiment #2......Page 72 6.4.3 AlexNet: Experiment #3......Page 73 6.5 Summary......Page 76 7 Training VGGNet on ImageNet......Page 77 7.1 Implementing VGGNet......Page 78 7.2 Training VGGNet......Page 83 7.3 Evaluating VGGNet......Page 87 7.4 VGGNet Experiments......Page 88 7.5 Summary......Page 90 8.1 Understanding GoogLeNet......Page 91 8.1.2 GoogLeNet Architecture......Page 92 8.1.3 Implementing GoogLeNet......Page 93 8.1.4 Training GoogLeNet......Page 97 8.3 GoogLeNet Experiments......Page 101 8.3.1 GoogLeNet: Experiment #1......Page 102 8.3.2 GoogLeNet: Experiment #2......Page 103 8.3.3 GoogLeNet: Experiment #3......Page 104 8.4 Summary......Page 105 9.1 Understanding ResNet......Page 107 9.2 Implementing ResNet......Page 108 9.3 Training ResNet......Page 114 9.5.2 ResNet: Experiment #2......Page 118 9.5.3 ResNet: Experiment #3......Page 119 9.6 Summary......Page 122 10.1.1 The Fire Module......Page 123 10.1.2 SqueezeNet Architecture......Page 125 10.1.3 Implementing SqueezeNet......Page 126 10.2 Training SqueezeNet......Page 130 10.4.1 SqueezeNet: Experiment #1......Page 134 10.4.2 SqueezeNet: Experiment #2......Page 136 10.4.3 SqueezeNet: Experiment #3......Page 137 10.4.4 SqueezeNet: Experiment #4......Page 138 10.5 Summary......Page 141 11.1.1 The FER13 Dataset......Page 143 11.1.2 Building the FER13 Dataset......Page 144 11.2 Implementing a VGG-like Network......Page 149 11.3 Training Our Facial Expression Recognizer......Page 152 11.3.2 EmotionVGGNet: Experiment #2......Page 155 11.3.3 EmotionVGGNet: Experiment #3......Page 156 11.3.4 EmotionVGGNet: Experiment #4......Page 157 11.4 Evaluating our Facial Expression Recognizer......Page 159 11.5 Emotion Detection in Real-time......Page 161 11.6 Summary......Page 165 12.1 The Indoor CVPR Dataset......Page 167 12.1.1 Building the Dataset......Page 168 12.2 Extracting Features......Page 172 12.3 Training an Orientation Correction Classifier......Page 175 12.4 Correcting Orientation......Page 177 12.5 Summary......Page 179 13.1 The Stanford Cars Dataset......Page 181 13.1.1 Building the Stanford Cars Dataset......Page 182 13.2 Fine-tuning VGG on the Stanford Cars Dataset......Page 189 13.2.1 VGG Fine-tuning: Experiment #1......Page 194 13.2.2 VGG Fine-tuning: Experiment #2......Page 195 13.2.3 VGG Fine-tuning: Experiment #3......Page 196 13.3 Evaluating our Vehicle Classifier......Page 197 13.4 Visualizing Vehicle Classification Results......Page 199 13.5 Summary......Page 203 14.1 The Ethics of Gender Identification in Machine Learning......Page 205 14.2 The Adience Dataset......Page 206 14.2.1 Building the Adience Dataset......Page 207 14.3 Implementing Our Network Architecture......Page 221 14.4 Measuring “One-off” Accuracy......Page 223 14.5 Training Our Age and Gender Predictor......Page 226 14.6 Evaluating Age and Gender Prediction......Page 229 14.7.1 Age Results......Page 232 14.7.2 Gender Results......Page 233 14.8 Visualizing Results......Page 235 14.8.1 Visualizing Results from Inside Adience......Page 236 14.8.2 Understanding Face Alignment......Page 240 14.8.3 Applying Age and Gender Prediction to Your Own Images......Page 242 14.9 Summary......Page 246 15.1 Object Detection and Deep Learning......Page 249 15.1.1 Measuring Object Detector Performance......Page 250 15.2.1 A Brief History of R-CNN......Page 252 15.2.2 The Base Network......Page 256 15.2.3 Anchors......Page 257 15.2.4 Region Proposal Network (RPN)......Page 259 15.2.5 Region of Interest (ROI) Pooling......Page 260 15.2.6 Region-based Convolutional Neural Network......Page 261 15.3 Summary......Page 262 16.1 The LISA Traffic Signs Dataset......Page 263 16.2 Installing the TensorFlow Object Detection API......Page 264 16.3.1 Project Directory Structure......Page 265 16.3.2 Configuration......Page 267 16.3.3 A TensorFlow Annotation Class......Page 269 16.3.4 Building the LISA + TensorFlow Dataset......Page 271 16.3.5 A Critical Pre-Training Step......Page 276 16.3.6 Configuring the Faster R-CNN......Page 277 16.3.7 Training the Faster R-CNN......Page 282 16.3.8 Suggestions When Working with the TFOD API......Page 284 16.3.10 Faster R-CNN on Images and Videos......Page 288 16.4 Summary......Page 292 17.1.1 Motivation......Page 295 17.1.2 Architecture......Page 296 17.1.3 MultiBox, Priors, and Fixed Priors......Page 297 17.1.4 Training Methods......Page 298 17.2 Summary......Page 299 18.1 The Vehicle Dataset......Page 301 18.2.1 Directory Structure and Configuration......Page 302 18.2.2 Building the Vehicle Dataset......Page 304 18.2.3 Training the SSD......Page 309 18.2.4 SSD Results......Page 312 18.2.5 Potential Problems and Limitations......Page 313 18.3 Summary......Page 314 19 Conclusions......Page 315 19.1 Where to Now?......Page 316