Home / Summaries / oefenen / object-bounding-boxes

Performance measures for object detection

5 important questions on Performance measures for object detection

Give two reasons why object classification is more complex than image classification.

1. Label has to be found, which relies on position and size.
2. Label has to be matched on its presence in the image.

What is the use of an overlapping threshold in evaluating the performance of an object detection?

It describes how much overlapping between the guessed bounding box and the actual bounding box of the respective object has to be in order to consider the region to be correct.

How is the algorithm called to compute the overlapping threshold between two bounding boxes? How to compute it?

Intersection over Union (IoU) = Area of Overlap / Area of Union

What is the algorithm called to filter out duplicated bounding boxes (that identify the same object) in object detection? How does it work?

Non-maximum suppression.
1. First it sorts the list of bounding boxes on their respective confidence score.
2. Iteratively remove bounding boxes that overlap with those of higher confidence score.

Name three performance measures for object detection.

1. Reduce to binary problem and apply F-Score.
2. False positive per window: Less false positives require less computation power.
3. False positive per image: Same as 2.