Lecture 12 - Detection

Question 1

What is object detection?

Answer 1

Object detection is the process of identifying and locating objects within an image, typically providing a bounding box and a class label for each detected object.

Question 2

Explain the sliding window approach in object detection.

Answer 2

The sliding window approach involves scanning the image with a fixed-size window at different scales and positions, applying a classifier to each window to detect objects.

Question 3

What is the Viola-Jones face detector?

Answer 3

The Viola-Jones face detector is a real-time object detection framework that uses integral images for fast feature evaluation, boosting for feature selection, and a cascade of classifiers for high detection rates.

Question 4

Describe the generalized Hough transform.

Answer 4

The generalized Hough transform is a method for detecting arbitrary shapes by mapping edge points in the image to a parameter space and identifying peaks in the accumulator array that correspond to shape instances.

Question 5

What are region-based methods in deep learning for object detection?

Answer 5

Region-based methods, such as R-CNN, Fast R-CNN, and Faster R-CNN, involve generating region proposals, extracting features from these regions, and classifying them to detect objects.

Question 6

Explain the YOLO (You Only Look Once) detection method.

Answer 6

YOLO is a real-time object detection method that divides the input image into a grid, with each grid cell predicting bounding boxes, confidence scores, and class probabilities for objects whose centers fall within the cell.

Question 7

What is a Feature Pyramid Network (FPN)?

Answer 7

An FPN is a deep learning architecture that creates a feature pyramid from a single input image, enabling object detection at multiple scales by combining features from different layers of a convolutional network.

Question 8

Describe instance segmentation.

Answer 8

Instance segmentation is a task that combines object detection and semantic segmentation, providing a pixel-wise mask for each detected object, distinguishing between different instances of the same class.

Question 9

What is pose estimation?

Answer 9

Pose estimation involves detecting and predicting the spatial configuration of an object’s key points, such as joints in a human body, often used for applications like action recognition and augmented reality.

Question 10

Explain 3D object detection.

Answer 10

3D object detection involves identifying and localizing objects in three-dimensional space, often providing 3D bounding boxes or poses, and is used in applications like autonomous driving and robotics.

Question 11

What is the purpose of non-maximum suppression in object detection?

Answer 11

Non-maximum suppression is used to remove redundant bounding boxes for the same object by selecting the box with the highest confidence score and discarding others that have a high overlap (IoU) with it.

Question 12

Describe the concept of anchor boxes in object detection.

Answer 12

Anchor boxes are predefined bounding boxes of different scales and aspect ratios used in region proposal networks (RPN) to detect objects at multiple scales and locations within an image.

Question 13

Write the formula for the confidence score in YOLO.

Answer 13

Question 14

Provide the formula for the weighted error in AdaBoost.

Answer 14

Question 15

What is the formula for updating weights in AdaBoost?

Answer 15

Question 16

How does the sliding window approach handle different object scales?

Answer 16

The sliding window approach handles different object scales by resizing the input image or the window itself, ensuring that objects of various sizes can be detected.

Question 17

What are integral images and how are they used in the Viola-Jones detector?

Answer 17

Integral images are data structures that allow for rapid computation of the sum of pixel values in rectangular regions. They are used in the Viola-Jones detector to quickly compute Haar-like features for object detection.

Question 18

Explain the role of boosting in the Viola-Jones face detector.

Answer 18

Boosting in the Viola-Jones face detector is used to select and combine weak classifiers (features) into a strong classifier, improving detection accuracy by focusing on the most informative features.

Question 19

How does the R-CNN method generate region proposals?

Answer 19

R-CNN generates region proposals using selective search, which combines hierarchical grouping of similar regions and exhaustive search methods to propose candidate bounding boxes for potential objects.

Question 20

What is the main advantage of Faster R-CNN over R-CNN and Fast R-CNN?

Answer 20

Faster R-CNN integrates region proposal generation into the network using a Region Proposal Network (RPN), significantly reducing computation time and enabling end-to-end training.

Question 21

Describe the YOLO detection pipeline.

Answer 21

The YOLO detection pipeline divides the input image into a grid, with each grid cell predicting bounding boxes, confidence scores, and class probabilities. Non-maximum suppression is then applied to remove redundant boxes.

Question 22

What is the role of the feature pyramid network (FPN) in object detection?

Answer 22

The FPN creates a pyramid of feature maps at different scales from a single input image, allowing the network to detect objects at multiple scales and improving detection accuracy for small objects.

Question 23

How does Mask R-CNN extend Faster R-CNN for instance segmentation?

Answer 23

Mask R-CNN extends Faster R-CNN by adding a branch that outputs a binary mask for each detected object, allowing for pixel-level segmentation in addition to bounding box detection.

Question 24

What are the main challenges in 3D object detection?

Answer 24

Main challenges in 3D object detection include handling occlusions, varying lighting conditions, and accurately estimating the object’s position, orientation, and scale in three-dimensional space.

Question 25

How does the generalized Hough transform handle arbitrary shapes?

Answer 25

The generalized Hough transform maps edge points in the image to a parameter space, where peaks in the accumulator array indicate the presence of shapes that match the model, allowing detection of arbitrary shapes.

Question 26

Explain the concept of region proposal networks (RPN) in Faster R-CNN.

Answer 26

RPNs generate region proposals by sliding a small network over the feature map, predicting bounding boxes and objectness scores at each location, integrating region proposal generation into the detection network.

Question 27

What is the purpose of non-maximum suppression in YOLO?

Answer 27

Non-maximum suppression in YOLO is used to eliminate redundant bounding boxes by selecting the box with the highest confidence score and discarding others with a high overlap (IoU), ensuring a single detection per object.