Which of the following is an operation not a task in computer vision

Object detection

Image convolution

Gradient computation

Max-pooling

Instance segmentation

Perspective projection

————————————————————————————————

921600

————————————————————————————————

It is more computationally efficient because segment sizes are bounded, limiting the number of pixels examined at each iteration

It limits distance between pixels by a certain threshold, utilizing the notion of hard spatial neighbourhood

It utilizes a more robust distance metric, rather than simple Euclidean distance used in k-means method

Unlike k-means, SLIC is a supervised learning method and thus can use labels to improve segmentation

————————————————————————————————

To undo channel concatenation by decreasing the number of convolutional feature maps

To undo convolution by applying the transposed convolution

To help backpropagate errors by introducing sparse convolutions

To undo pooling by outputting an image with larger resolution (i.e., pixels in spatial directions)

————————————————————————————————

We output maximal values at their respective indexes (called max location switches) and place zeroes elsewhere

We do bilinear interpolation to compute the output

We use ‘bed of nails’: output the maximal values in the top left corner and zeros elsewhere

————————————————————————————————

A matrix produced by computing dot product between two sets of vectors

A matrix of feature activations in a CNN

A confusion matrix of CNN

A positive-semidefinite matrix used to generate random numbers from a Gaussian distribution

————————————————————————————————

It achieves superior performance in generating Gaussian mixtures

It produces data that is hard to distinguish from real

It produces nicely looking images

Week 5 >>> Which of the following is an operation not a task in computer vision >>> Deep Learning in Computer Vision