Venue Student Number Family Name First Name ____________________ |__|__|__|__|__|__|__|__| _____________________ _____________________ School of Electrical Engineering, Computing and Mathematical Sciences Mid-Semester Test Test Duration Reading Time Semester 2, 2019 COMP3007 Machine Perception This paper is for Bentley Campus students 1 hour 10 minutes Students may write notes in the margins of the test paper during reading time Total Marks 100 Aids to be supplied by the University: None Aids to be supplied by the Student: None Calculator: No calculators are permitted in this test Conditions This is a CLOSED BOOK test no text books or written materials permitted. Mobile phones or any other device capable of communicating information are prohibited from use during tests. Use of any device capable of storing text or other restricted information is also prohibited (e.g. electronic organisers, PDAs, calculators with the capacity to store text information). Any breaches of this policy will be considered cheating and appropriate action will be taken as per University policy. Instructions to Students Answer all questions in the space provided in the quest paper. This test consists of 5 questions. Attempt ALL questions. Provide answers in the space below each question. Use a black/blue pen to write your answers. Answers written in pencil will be ignored. Mid-Semester Test, Semester 2, 2019 COMP3007 Machine Perception Question 1 Perception Pipeline (20 Marks) You are developing a machine vision system for automatic shark detection along the WA coastline. The developed vision algorithms should be able to process images from aerial cameras, detect the area of sharks presence and recognise sharks in images. Propose two suitable platforms where the developed machine vision system can be used for this task. Briefly describe their advantages and disadvantages. (8 marks) Sketch the pipeline of the proposed automatic shark detection system. Clearly indicate in your drawing the data flowing between the blocks and the final outputs from this pipeline. For each block in the pipeline, identify one or more challenges. (12 marks) Answer: Page 1 of 8 Question 2 - Image Processing (20 Marks) Question 2a Image Representation Suppose the intensity levels of images are represented with 8-bit numbers (0-255). How many bits do we need to store a 100x100 colour image? Answer: Question 2b Histogram Equalization With an example, explain how histogram equalization can be used to increase the contrast of an image. Answer: Page 2 of 8 Mid-Semester Test, Semester 2, 2019 COMP3007 Machine Perception [4 Marks] [8 Marks] Question 2c - 2D Transformation With the help of diagrams, explain what property is preserved in each of the following 2D transformations: 1) affine transformation, and 2) perspective transformation. [8 Marks] Answer: Page 3 of 8 Mid-Semester Test, Semester 2, 2019 COMP3007 Machine Perception Question 3 - Feature Detection (20 Marks) Points, edges, lines and regions are the common features to be detected in computer vision. a) Briefly explain the purpose to detect each of these four features, and for each of them, name a suitable application. [8 Marks] b) Edgedetectorsarebasedonimagegradients.Explainbrieflyhowtocomputeimage gradients using image convolutions and how many convolution kernels are required to compute image gradients. Maximally Stable Extremal Regions Algorithm [12 Marks] Page 4 of 8 Mid-Semester Test, Semester 2, 2019 COMP3007 Machine Perception Mid-Semester Test, Semester 2, 2019 COMP3007 Machine Perception Question 4 - Feature Extraction (20 Marks) Question 4a - Binary shape analysis In the binary image shown below, the foreground pixels of interest are the white boxes. Suppose that the objects are 4-connected and you are using a two-pass connected- component labelling algorithm to find and label the blobs. 1. Indicate the labels obtained after the first and second passes; and 2. Provide the sets of equivalent labels after the first pass; and 3. State the final number of blobs detected and compute the area of each blob. Answer: [15 marks] Figure 1: First pass. Figure 2: Second pass. Page 5 of 8 Question 4b HOG and SIFT In 3D scene reconstruction from multiple 2D images which are taken from different views of the same 3D scene, key point matching between images are required. Which feature descriptors, HOG or SIFT, are more suitable? Justify your answer. [5 marks] Page 6 of 8 Mid-Semester Test, Semester 2, 2019 COMP3007 Machine Perception Question 5 - Clustering (20 Marks) The patterns in Figure 3 include two clusters which are depicted by diamonds and circles respectively, and two outliers which are depicted by triangles. All the grids have the same size which is 1cm x 1cm. Explain the advantages of DBSCAN when compared with k-Means, and use the example in Figure 3 to show these advantages. Clearly show that it is not possible for k-Means to simultaneously cluster diamond points into one cluster and cluster those circle points into another cluster, and show that these points can be clustered into two clusters by DBSCAN when proper density parameters, namely the neighbourhood radius and the number of points that define a threshold for core points, are selected. At least two advantages of DBSCAN should be discussed. Figure 3 Page 7 of 8 Mid-Semester Test, Semester 2, 2019 COMP3007 Machine Perception Mid-Semester Test, Semester 2, 2019 COMP3007 Machine Perception This page is intended to be blank. END OF TEST Page 8 of 8