Tasks:
The visual tasks could range from instance recognition to human action recognition. In
instance recognition, we would be answering specific visual identification questions such as: is this
an Airbus A380? Another relevant question is object classification, where we aim to answer
questions such as: does this image contain a bike or not? Another relevant task is that of object
detection in images and videos: where is the bike in the image? In action recognition we aim at
more general tasks such as: what is going on in the video? In the course we will undertake a study
of different tasks.
Techniques:
In terms of techniques, there have been a wide range of machine learning techniques
ranging from Adaboost and support vector machines to state of the art deep learning techniques.
Many of the machine learning techniques have attained popularity based on their success in visual
recognition tasks. Indeed, the success of adaboost for face detection has made boosting popular
while deep learning techniques became widely popular once they succeeded in large scale object
classification. In this course we aim to understand a few of the machine learning techniques
involved as applied to visual recognition.
Advances:
There have been certain assumptions in visual recognition such as the need for large
number of manually supervised training samples. While this has been dominant there are a number
of techniques that aim to relax this assumption by minimising the need for supervision. These
include learning with latent variables, active learning techniques, unsupervised machine learning
techniques. In the final part of the course we aim to study these advanced techniques.
A brief outline of the topics to be covered in the course are as follows: