QUIZ

1. A 2D rotation transform requires two parameters to be fully specified

2. SIFT is a rotation invariant descriptor

If we initialize the k-means clustering algorithm with the same number of clusters but different starting positions for the centers, the algorithm will always converge to the same solution

4. Corners are the only type of keypoints that can be used in images

3. SIFT uses histogram of image intensities as a descriptor

Given a set of 3 images as shown in next figure, finding and labeling the image in the center as ”containing cat” is considered a detection task in recognition

We need labels when performing clustering

9. When clustering an image using color values, regions belonging to same cluster needs to be adjacent

We need class labels when performing regression

11. Viola Jones can only be used for training a frontal face detector

10. K means will always result in the clusters independent of initialization of cluster centers

Being a template based method, Viola Jones can be also very suitable for deformable objects

In the K-means clustering algorithm, after convergence the distortion is minimized

13. Features extracted from images for the purpose of image matching need not to be distinctive

Given an image containing a cat, automatically finding the contour of the cat in the image is considered an object detection task.

One can make a visual descriptor rotationally invariant by assigning orientations to the key points and then rotating the patch to a canonical orientation. In SIFT this is done by constructing Histograms of Gradients in a neighborhood around each pixel in the image and assigning the largest bin as the corresponding direction of the keypoint

A Gaussian pyramid is a hierarchy of images. The bottom layer is an input image. The next layer is obtained by blurring the image in the previous layer and upsampling it, and so on. They are used in many computer vision algorithms such as SIFT

18. Estimating the price of a house is an example of a classification problem

In general we need more manual work when building a labelled dataset for semantic image segmentation compared to the case of image classification

Assume that while training a machine learning the training loss at an iteration was L = 90 and the weight vector was w=[■(1&3&-10)]^T. At a later iteration the loss became L = 87 and w=[■(1&-5&10)]^T. Consider we are using a validation set for which the validation losses at the two iterations are 150 and 160 respectively, the second weight vector would be preferred from a validation point of view

Assume that while training a machine learning the training loss at an iteration was L = 90 and the weight vector was w=[■(1&3&-10)]^T. At a later iteration the loss became L = 87 and w=[■(1&-5&10)]^T. From a training point of view the first weight vector is better

22. When we have less data and the model complexity stays the same, overfitting is more likely.

26. It is always better to use a linear classifier as opposed to a more complex classifier

27. In building an ensemble classifier it is always good to have classifiers that are dependent in terms of their classification outputs or scores

30. Assume the validation loss at successive iterations of the gradient descent method is as follows: 100, 90, 82, 80, 65, 70, 50

29. In a well-trained convolutional neural networks for classifying images the first few layers correspond to filters capturing bigger parts of the objects trained on while later layers represents simpler patterns such as edges and corners

31. If we use a simple stopping criterion to stop the training then the algorithm should stop when it sees a validation of 70

32. A good strategy in initializing weights in a neural network is to initialize them all as zeros

Bag of words feature matching has a main advantage of taking into account spatial feature arrangement when matching images

36. Given an image containing a cat, automatically labelling the image as ”containing cat” (without specifying where the cat exists in the image) is considered an object detection task

35. Bag of words feature matching does not take into account spatial feature arrangement when matching images

37. One can make a visual descriptor rotationally invariant by assigning orientations to the key points and then rotating the patch to a canonical orientation. In SIFT this is done by constructing Histograms of Gradients in a neighborhood around the feature point and assigning the largest bin as the corresponding direction of the keypoint. Later, all detected features are rotated so that the corresponding orientations are vertically aligned

In regular shallow neural networks one needs to use a non-linear activation function between layers while in deep neural networks this is not needed

40. In regression, the predicted output is discrete, this is in contrast to classification

38. One can make a visual descriptor rotationally invariant by assigning orientations to the key points and then rotating the patch to a canonical orientation. In SIFT this is done by constructing Histograms of Gradients in a neighborhood around each pixel in the image and assigning the largest bin as the corresponding direction of the keypoint. Later, all detected features are rotated so that the corresponding orientations are vertically aligned.

In mean shift clustering algorithm we find cluster centers as the modes of the probability density function of input features

42. Object localization is different than object detection because in object localization we could have multiple instances of the target object

45. Some of the reasons we use convnets instead of fully connected layers in images are computational efficiency (sharing of weights) and preserving spatial structure

We can solve the object detection problem as a regression problem to find all boxes of objects in images

The main idea behind resnet blocks is to add shortcuts in the network blocks so that we can learn identity mapping in the worst case scenario. This helps learning even when we increase number of layers

48. In the case of L2 regularization we have two terms in the loss function: a data term and a regularization term. The regularization term is multiplied by a parameter, let’s call it “lambda”. If we increase lambda we are risking more overfitting

50. Single shot detection (SSD) and You-only-look-once (YOLO) detection algorithms operate on regions extracted from a separate region proposal network (separate from the network that performs detection

53. The goal of the L2 loss in the following architecture is to predict the location of the object of interest

In designing cascaded object detection systems it is usually the case that the first detector in the cascade has high recall and possibly high false positive rate

56. Random forest classifier is an example of an ensemble classifier methods

If we want to build a neural network estimating 4 joint positions on an image, one possible way is to formulate it as a regression problem of x,y locations of the joints and have an 8 output unit network

57. Running a detector through an image results in many possible overlapping boxes with different detection scores, a popular method to select the best detection window(s) is called Non-maxima suppression

A condition for a point to be a corner feature in an 2-D image is that the two eigen values of the M matrix (of gradients) are small

60. Using gradient-based feature descriptors guarantees invariance against linear illumination changes

Estimating the house of a price is an example of a classification problem

61. The goal of the L2 loss in the following architecture is to predict the location of the object and the object type

In SSD/Yolo object detector we reply on a region proposal method/network

65. A condition for a point to be a corner feature in an 2-D image is that there is a large intensity gradient along the X, Y and depth directions of the image

The idea of the inception block is to test various sizes of convolution filters

Using gradient-based feature descriptors guarantees invariance against non-linear illumination changes

The goal of the L2 loss in the following architecture is to predict the location of the object of interest only

71. In the basic gradient descent optimization algorithm, we sample a batch of images in each iteration and update the weights accordingly

70. The main idea behind Resnet blocks is to add shortcuts in the network blocks so that we can learn identity mapping in the worst-case scenario. This helps learning even when we increase the number of layers.

Suppose a dataset is linearly separable, a logistic regressor with regularization parameter λ > 0 is guaranteed to separate the data