Filter löschen
Filter löschen

How to implement Faster R CNN object detector?

1 Ansicht (letzte 30 Tage)
Jay am 20 Jun. 2017
Beantwortet: Birju Patel am 20 Jul. 2017
I am using
for the detection task. My ROIs in an image has single size((96*96*3) in (684*912) image size.) Although my object sizes very in a 96*96*3 image patch, I train my CNN such that it classifies the objects at the centres only.
In such a scenario, I want the anchors in RPN such that it detects foreground even for the smallest object at the centre(let's say objects with 60 diameters) so that it could be the part of the classification (It actually helps when two objects are overlapping each other, and one of them is slight off-center, I believe).
In such case I need might need three different anchorBox sizes(let's say 60,80,96). For such case how to decide
'BoxPyramidScale' Anchor box pyramid scale
'NumBoxPyramidLevels' Number of anchor box pyramid levels
I am quite confused in this two arguments. Thanks.

Antworten (1)

Birju Patel
Birju Patel am 20 Jul. 2017
Hi Jay,
Because you have objects that have a single size, 96x96, you can set BoxPyramidScale to 1 and NumBoxPyramidLevels to 1.
BoxPyramidScale and NumBoxPyramidLevels control how the RPN anchor boxes are scaled up from the MinBoxSizes. These two parameters define a "box pyramid" where the MinBoxSizes define the smallest RPN anchor boxes, and each level of the pyramid contains scaled versions the boxes.
This lets you generate RPN anchors of different sizes, which enables Faster R-CNN to detect objects of different sizes.
HTH, Birju

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by