代写ImageNet留学生、VGG network代做、代写C++/Python代写、代做Java程序

日期：2019-04-10 10:06

Deep Learning

Which statements are correct? Circle the item number of all correct answers.

1. Yann LeCun won the 2018 Turing Award for conceptual and engineering breakthroughs that have made deep

neural networks a critical component of computing.

2. Yann LeCun applied the back-propagation algorithm to a multi-layer neural network to train it to recognize

handwritten zipcode digits provided by the U.S. Postal Service.

3. NETtalk is a historic multi-layer neural network that learned to read English aloud.

4. The convolution of a row of image values (0, ... 0,0,1,1,1,0,0,..., 0) with itself is (0, ..., 0, 0, 2, 2, 2, 0, 0, ..., 0 ).

5. The convolution of a row of image values (0, ... 0,0,1,1,1,0,0,..., 0) with itself is (0, ..., 0, 0, 1, 2, 3, 2, 1, 0, 0, ...,

0 ).

6. Max pooling with a 4x4 filter and a stride 1 results in a two row matrix where the values of each row are (8,8,10).

7. To train a deep network, AI system designers use the gradient descent update rule with respect to a fraction of

training examples to adjust the network weights during the back propagation algorithm.

8. The term ’one-hot encoding’ when used with a deep neural net means that the output vector has zeros for all

entries except one value is 1.

9. The dropout trick removes under-performing nodes from the network.

10. In a deep neural network, the softmax function is applied to nodes in the last layer so that the outputs can be

interpreted as probabilities.

11. The Alex Net uses sigmoid functions as activation functions.

12. In the 2012 ImageNet Challenge, the Alex Net performed well on object classification, but the VGG network

was better at localizing large objects.

13. Distractor images make the problem of face verification more difficult.

2

Computer Vision and Geometry

(a) Use the perspective projection equations we discussed in class to solve the following problem:

A scene point with 3D world coordinates (500, 600, 1000) is projected in a pinhole camera at coordinates (25, 30),

where both are in millimeters in the camera’s reference frame and the image coordinates have their origin at the

camera’s principal point. What is the focal length f of the camera? Sketch the geometry and show your calculations.

f =

(b) Use the binocular stereo equation we discussed in class to solve the following problem: A scene point is projected

in two pinhole cameras with parallel optical axes and a baseline of 10 cm. The focal length of both cameras is 25 mm,

and the width of a pixel in each camera is 1.4 micrometer. The scene point is projected with a disparity of 30 pixels.

What is the depth of the scene point in meters? Sketch the geometry and show your calculations.

Z =

3

Interpreting Camera Motion

Assume the following computer vision scenario:

Fixed environment

Fixed lighting

The origin of the word coordinate system is the center of projection of the camera.

The image plane in the 3D world coordinate system is located at Z = 1 with the X- and Y-axes of the world

coordinate system parallel to the x- and y-axes of the image plane

The optical axis runs along the Z-axis piercing the image plane at the point (0, 0, 1)T

The real world point R = (X, Y, Z)T is related to the image point r = (x, y, 1)T by the perspective projection

equation.

A camera is moving with rotational velocity w = (A, B, C)T and translational velocity t = (U, V, W)T . The flow

field can be described by

where (X, Y, Z)T describes the scene coordinates and (x, y)T the respective image coordinates.

(a) Assume you measure the optical flow field shown below. Give an expression of the translational velocity (u, v)T

of the camera as a function of the parameters U, V, and/or W.

(u, v)T =

(b) Give an expression of (u, v)T as a function of the parameters U, V, W, A, B, and/or C for each of the special

cases below. Simplify the expression by taking into account which of the parameters is zero. Draw examples of the

respective flow fields by selecting values for the parameters.

4

(i) Camera A moves forward along the optical axis. (u, v)T =

(ii) Camera B moves backwards along the optical axis but twice as fast as camera A. (u, v)T =

(iii) Camera C rotates counter-clockwise around the optical axis. (u, v)T =

5

(c) If the following two consecutive images are captured by the moving camera, what can you say about the direction

of the movement of the camera?

(d) The Lukas-Kanade and Horn-Schunk algorithms to estimate the optical flow in an image sequence both rely on the

Constant Brightness Assumption.

Circle the correct answer:

True or False?

(e) Motion segmentation is the task of explaining a flow field that is due to the movement of different objects and/or

the camera.

Circle the correct answer:

True or False?

6

Hidden Markov Models

In the network below, assume ?1 = 0.5, ?2 = 0.2, and ?3 = 0.3.

(a) What is the probability that the hidden Markov model observes the sequence V3V2

(b) What is the probability that the hidden Markov model observes the sequence V2V3?

(c) What is the probability that the hidden Markov model observes the sequence V1V3V3?

7