Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit

EXAMINATION

End of Semester 2, 2019

COMP3007 Machine Perception

Question 1 – Perception Pipeline (10 Marks)

Question 2 – Feature Extraction (10 Marks)

Describe the advantages and disadvantages of HOG, SIFT and CNN features. For each of these three image features, name one application for which this feature is more suitable than the others.

Question 3 – MSER (10 Marks)

What is Maximally Stable Extremal Region (MSER) used for? Describe the main steps of MSER and explain why the detected regions are called “maximally stable extremal regions”.

Question 4 – Object Localization (10 Marks)

Describe how to use sliding window methods to localize objects in images, and explain the conditions on the object model so that the Kadane's algorithm can be used to speed up the sliding window methods for object localization.

Question 6 – Deep Neural Networks (15 Marks)

Residual networks are the backbones of many state-of-art computer vision and machine learning algorithms.

· Briefly describe at least four key components of a residual network and explain why they are useful. (8 marks)

· Which property of the objective functions of residual networks are useful for parallel processing with GPUs? Explain how this property is used for parallel processing? (3 marks)

· Explain why residual networks perform better than AlexNet and VGGNet for image classification. Clearly identify the critical component of the residual networks and how this component is used to overcome the limitations of AlexNet and VGGNet. (4 marks)

Question 7 – Training of Classifiers (10 Marks)

Question 7a: In the training of many classification methods, hyper-parameters such as the regularization numbers of Support Vector Machines are involved. What is the risk of tuning the hyper-parameters on the test dataset? How hyper-parameters should be tuned to avoid this risk? [3 marks]

Question 7b: Dropout is often used in training deep neural networks. What is the purpose of using dropout? Is dropout used to speed up training? [3 marks]

Question 7c: What happens when the learning rate is too small in training deep neural networks? What happens when the learning rate is too large? Explain why larger learning rates are often used in the beginning of the training and smaller learning rates are used in the late stages of training [4 marks]

Question 8 - State Estimation (10 Marks)

You are developing a Kalman filter to estimate the position of a self-driving car which is assumed to run straight at approximately constant speed. A sensor is equipped to provide the position of the car on the road.

· Identify the states for the state transition model. (2 marks)

· Identify the sources of the process noise and the measurement noise. (2 marks)

· What conditions on the noises and the state transition models should be satisfied for Kalman filters to work successfully. (3 marks)

· Explain why both the measurement and the state transition model should be used together to estimate the position of the car. (3 marks)

Question 9b: The Kadane’s algorithm is an efficient algorithm which can find the largest- sum subarray of an array of numbers. You are asked to apply the Kadane’s algorithm to search for the string S = AAA from a text such as T = AAbcAAAdAeAf. Only one instance of the string S = AAA is required to be found if multiple such strings exist in the text. Explain how to transform the string search problem so that the Kadane’s algorithm is applicable for this task. (9 marks)