Week 1, Lecture 1: Course Introduction and Overview

Week 1, Lecture 1: Course Introduction and Overview#

NPTEL Deep Learning for Computer Vision | Prof. Vineeth N Balasubramanian, IIT Hyderabad

Companion notebook for §1.1 Introduction.

What you will learn

What computer vision is, and the definitions the field has given itself over forty years.
Where it is used, from autonomous vehicles and medical imaging to agriculture and remote sensing.
Why the problem is hard: inverse problems, high dimensionality, and the gap between perception and cognition.
What this course covers and how it is structured.

Contents

What is computer vision?
Why: applications
Why is it hard?
This course

No code in this lecture. This is a discussion lecture, so there is nothing to run here. The page follows the slides and ends with questions to think about. Hands-on notebooks begin at Lecture 4.

1. What is computer vision?#

A field that seeks to automate and endow a computing framework with the ability to interpret images the way humans do. A sub-topic of Artificial Intelligence.

Look at an ordinary photograph and you can answer, effortlessly, questions like where is the gluestick?, what is the full title of that book?, or what is wrong with this image? The question this field asks is whether a machine can answer them too.

Other definitions from the literature, each emphasising something different:

Definition	Source
“the construction of explicit, meaningful descriptions of physical objects from images”	Ballard and Brown, 1982
“computing properties of the 3D world from one or more digital images”	Trucco and Verri, 1998
“to make useful decisions about real physical objects and scenes based on sensed images”	Sockman and Shapiro, 2001

Notice the progression: from describing objects, to recovering the 3D world, to making decisions. The field’s sense of what counts as “vision” has widened over time.

2. Why: applications#

Computer vision now sits inside a very wide range of systems:

Autonomous vehicles and driver assistance
Medical imaging and diagnosis
Surveillance and security
Factory automation and robotics
Human-computer interaction
Visual effects in film
Retail (cashier-less stores, virtual try-on, shrinkage detection)
Healthcare (blood loss estimation, dermatology screening)
Document understanding (OCR, robotic process automation)
Agriculture (crop sensing, livestock identification)
Banking and finance (mobile cheque deposit, insurance risk profiling)
Telecom and social media (image understanding, brand exposure analytics)
Augmented reality (visual support, warehouse management)
Remote sensing (land use, forestry modelling)
Structural health monitoring (oilwell inspection, drone-based bridge inspection and 3D reconstruction)

3. Why is it hard?#

Vision is an inverse problem. Physics and computer graphics work forward: given geometry, materials, lighting and a camera, produce an image. Vision has to run that backwards. From the image alone, with no knowledge of how the picture was taken or what the camera settings were, we must recover shape, lighting, colour, objects and their interactions. That inference is made from incomplete, partial and noisy information, and in general it has no unique answer.

The data is enormous. Images are extremely high-dimensional, which makes the computation heavy before any cleverness is added.

Perception is not the same as measurement. The human visual system does not report pixel values, and a long list of illusions makes the point:

the Müller-Lyer illusion: two lines of equal length look different,
a variation of the Hermann grid: spots appear at intersections that are not there,
Adelson’s brightness constancy illusion: two patches of identical intensity look very different,
counting the red Xs in two arrangements: identical content, very different difficulty.

And we do not have a complete model to copy. No full model of the human visual system exists; the models we do have describe subsystems rather than the whole. Which parts of a scene are perceived and which are cognized? When does an object matter for a task and when does the context matter more? Computer vision is, in this sense, AI-complete: solving it in general looks equivalent to solving artificial intelligence.

4. This course#

The course is organised in three broad movements:

Classical vision (Weeks 1 to 2): how images are formed and represented, and the hand-designed features (edges, blobs, corners, SIFT, HOG) that dominated the field before deep learning.
Deep learning for vision (Weeks 3 to 6): neural networks and their training, convolutional networks, the architectures that made them work, and their use for detection and segmentation.
Modern topics (Weeks 7 to 12): recurrent models, attention and transformers, vision transformers, generative models including GANs, VAEs and diffusion, and vision-language models.

Each week pairs lecture videos with slides, and from Lecture 4 onwards with a hands-on notebook you can run in Colab.

Key takeaways#

Idea	Takeaway
What it is	Automating the interpretation of images the way humans do; a sub-field of AI
How definitions shifted	From describing objects, to recovering the 3D world, to making decisions
Where it is used	Across transport, medicine, retail, agriculture, finance, media and infrastructure
Why it is hard	It is an inverse problem, solved from partial and noisy data, on very high-dimensional inputs
Perception is not measurement	Illusions show the visual system reports something other than raw intensities
AI-complete	General vision appears to require general intelligence

Exercises#

Pick an illusion. Take the Adelson brightness constancy illusion and measure the actual pixel intensities of the two patches. They are identical. Why does a system that reports the correct numbers still fail to describe what a human sees?
Forward versus inverse. Write down what a graphics renderer is given and what it produces. Now reverse the arrows. Which of the inputs can be recovered uniquely from a single image, and which cannot?
Choose an application from the list above and write down the specific visual question the system must answer, the failure that would be most costly, and whether an error is recoverable.
AI-complete? Argue for or against the claim that general computer vision requires general intelligence. What would a counterexample look like?