Understanding the positive and negative overlap range

Hi all and thank you for responding to my questions in advance!

I am trying to obtain a simple understanding of the negative and postive ranges.

I read the documentation in matalb for the understanding but i still don't get it and the explanation there is still complex!

% Adjust NegativeOverlapRange and PositiveOverlapRange to ensure
% that training samples tightly overlap with ground truth

'PositiveOverlapRange' A two-element vector that specifies a range of
%                        bounding box overlap ratios between 0 and 1.
%                        Region proposals that overlap with ground truth
%                        bounding boxes within the specified range are used
%                        as positive training samples.
%
                        Default: [0.5 1]
%
'NegativeOverlapRange' A two-element vector that specifies a range of
%                        bounding box overlap ratios between 0 and 1.
%                        Region proposals that overlap with ground truth
%                        bounding boxes within the specified range are used
%                        as negative training samples.
%
                        Default: [0.1 0.5]

I am aware of what 3 variables after the trainRCNNObjectDetector are and what they do and how to achieve this! but ranges are confusing me understanding!

my questions in regards to image processing;

1. what is the threshold actually controlling/ doing for the positive and negative overlap range
2. Is there a link to understand this on youtube etc to get a simple break down of what this does or is? I have been trying this but maybe my terminology is incrorrect!!
3. I specified only the negative range, what happends when I don't specify the positive range?
4. what happends when i specify both positive and negative ranges?
5. what am I really telling the system to do actually?!!!?!?!!!?!
6. if I modify the Positive Overlap Range, What am I Actually Doing, Same for the Negative Over Lap Range?

I have my code taken from the rcnn stop sign example in math lab;

rcnn = trainRCNNObjectDetector(BCombineData, Tlayers, options, 'NegativeOverlapRange', [0 0.3]);
rcnn = trainRCNNObjectDetector(BCombineData, Tlayers, options, 'PositiveOverlapRange', [0.5 1] ,'NegativeOverlapRange', [0 0.3]);
rcnn.RegionProposalFcn;
network = rcnn.Network;
layers = network.Layers;

[Matpar]

thank you for elaborating on this!

Answer 1

[Harsha Priya Daggubati]

Hi,

As you might know, Object Detection involves dividing the input image into multiple pieces and identifying the presence of object in each individual piece. An enhancement to this is involving segmentation process in Object Detection using Region Proposal Networks in conjunction with Fast RCNN Algorithm.

Answering your queries:

1. Positive/Negative Overlap Range specify to the nework to treat the region under consideration as positive/negative (in the presence of an object), by computing Intersection over Union (IoU) with ground truth data.
2. I guess reading about Object Detectors (CNNs and RCNNS) and their working would help you.
3. If specified only Negative Range, Positive Overlap Range will be assigned a default value of [0.5 1].
4. The specified ranges will be considered.
5. You are giving a criterion to the network based on which it classifies whether the object of importance is present in the image under consideration or not.
6. You are adjusting your networks ability to match with the ground truth data.

Hope this helps!