Video Analysis

————————————————————————————————

Correlation between two vectors

Average Precision

Endpoint Error

Detection Error Tradeoff curve

Angular Error

————————————————————————————————

The number of feature maps required to produce an appropriate robustness for the tracker

The number of convolutions required to achieve a specified tracking quality

The time required for tracking algorithm to run compared to the time required for image filtering operation to run

————————————————————————————————

False negative error

False positive error

False acceptance error

ID switch

Mean absolute error

False coverage error

————————————————————————————————

0.8

————————————————————————————————

To localize actions in videos we usually detect and track relevant objects first, and then apply action classification in a temporal window along the track.

It is easy for convolutional neural network to extract and use motion information automatically, when applied to whole video volume.

In dense trajectories with CNN features, point neighbourhoods are cropped from frames along the trajectory, concatenated into space-time volume along the trajectory, and then supplied to CNN for feature computation.

By explicit consideration of motion information in form of optical flow maps, point and keypoint trajectories, we can currently improve the performance of action recognition.

[rank_math_breadcrumb]

Week 4 >>> Video Analysis >>> Deep Learning in Computer Vision

1.

2.

3.

4.

5.

6.

7.

8.

Object Detection

Why is the idea of having 1x1 convolutions reasonable

Which of the following is an operation not a task in computer vision

Basic image processing