Computer Vision Projects with PyTorch

Table of Contents

1. Frontmatter

2. Chapter 1. The Building Blocks of Computer Vision

   Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

   Abstract

   Humans have been part of a natural evolutionary pattern for centuries. According to the Flynn Effect, an average person born in recent times has a higher IQ than the average person born in the previous century. Human intelligence allows us to learn, decide, and make new decisions based on our learnings. We use the IQ score to quantify human intelligence, but what about machines? Machines are also part of this evolutionary journey. How have we moved our focus to machines and made them intelligent, as we know them today? Let’s take a quick look at this history.

3. Chapter 2. Image Classification

   Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

   Abstract

   The last chapter discussed several important concepts in computer vision. A few of the best practices in the field of computer vision were discussed as well, so it is time to practice those. This chapter sets the tone for multiple tasks in the field of computer vision. We start with a basic explanation of how to start using the Torch components to build a model, define a loss function, and train.

4. Chapter 3. Building an Object Detection Model

   Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

   Abstract

   Object detection is one of the most sought-after skills these days. An image can have multiple classes. In addition, classifying an object solves just part of the problem. The other part lies in the localization of the object. Object detection helps identify the class location of an image with a bounding box. The bounding box can be further processed for various sub-tasks. As an example, think about what a traffic cam needs to detect and identify cars.

5. Chapter 4. Building an Image Segmentation Model

   Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

   Abstract

   Images around us come in different textures, patterns, shapes, and sizes. They carry an enormous amount of information which can easily be understood by the human eye and brain, but is still a difficult problem for computers. Image segmentation is a problem set wherein we try to train computers to understand images so that they can separate dissimilar objects and unite similar objects. This can be in the form of similar pixel intensities or similar textures and shapes.

6. Chapter 5. Image-Based Search and Recommendation System

   Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

   Abstract

   In order to retain and acquire new customers, especially in the e-commerce arena, customer service needs to be top-notch. There are already thousands of e-commerce platforms and the number will only increase in the future. Platforms with excellent customer experience will survive in long term.

7. Chapter 6. Pose Estimation

   Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

   Abstract

   Human pose estimation (HPE) is a computer vision task that detects human poses by estimating major keypoints, such as eyes, ears, hands, and legs, in a given frame/video. Figure 6-1 shows an example of human pose estimation in action.

8. Chapter 7. Image Anomaly Detection

   Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

   Abstract

   The study of machine learning has put us in the course of studying various patterns and behavior. It has allowed us to build models that can study closed environments. Predictive power often follows the model training process. It is an important question that we need to ask often when we are training a model. There is another question that needs an answer—how much data is sufficient to help the model understand the distribution such that we can have a good representation? This chapter will work out an example and the concepts regarding these important questions. We are discussing anomaly detection in computer vision.

9. Chapter 8. Image Super-Resolution

   Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

   Abstract

   With the advent of high-resolution image capturing agents, the information captured in images is huge. Technology has moved from ultra HD to 4K and 8K resolutions. Movies are using high-resolution frames these days; however, there are also situations when they need to enhance a low-resolution image to a high-resolution one. Imagine a scene where the protagonist of a movie is trying to determine the license plate captured from a picture of a speeding car. Super-resolution can now help us zoom into an image to a high degree without distorting it. A few interesting advancements have happened in the industry and we are going to discuss those with some examples.

10. Chapter 9. Video Analytics

    Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

    Abstract

    The machine learning journey started from structured data long ago to the process of extracting meaningful predictions. As data grew, machine learning started exploring other data types as well. Today, there is no limit to the types of data that can be processed.

11. Chapter 10. Explainable AI for Computer Vision

    Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

    Abstract

    Most machine learning and deep learning models lack a way of explaining and interpreting results. Due to the dynamic nature of deep learning models and increasing state-of-the-art models, the current model evaluation is based on accuracy scores. This makes machine learning and deep learning black-box models. This leads to lack of confidence in applying the model and lack of trust of the generated results. There are multiple libraries that help us explain models of structured data like SHAP and LIME. This chapter explains computer vision model outputs.

12. Backmatter