Object detection

Question 1

Name some difficulties associated with object detection in images

Answer 1

Objects may appear in different poses
Cluttered scenes with occlusions
Intrinsic variability within a class
Non-rigid objects like humans

Question 2

What are the steps in Dalal-triggs pedestrian detection algorithm

Answer 2

1. Extract windows of 64x128 in all relevant positions and scales.
2. Compute HoG
3. Use support vector machine to score the window
4. Remove non-max overlapping detections

Question 3

Name some important properties of the non-linear functions used in deep learning

Answer 3

Good gradients,

efficient

Question 4

What is the field of vision regarding Convo nets?

Answer 4

The field of vision for a node is the input nodes that affect the given node in any way.

Question 5

What is the Relu

Answer 5

A function, f(x) = max(0,x)

Question 6

What might be the problem if a neural net doesn’t train?

Answer 6

1. If the input to the layer (The gradient) is zero, that layer won’t train
2. If the weights to a layer are zero, no layer before the layer will change.
3. If many layers have weights with norm smaller than one, the change will get smaller and smaller

Question 7

What solutions might be implemented in a neural net to resolve the problem of diminishing training as we go back through the layers?

Answer 7

1. Batch normalization
2. ResNet/ Skip connections
3. Good weight initialization

Question 8

Why is the softmax activation function often used in the last layer for classification problems?

Answer 8

The sum of all outputs will be equal to 1 with the softmax, which means the output values can be interpreted as “probabilities” for an object belonging to each class.