Fix for boundig box scaling issue.

[Pandoro]

This fixes a critical error where the bounding box was limited to 159x159 pixels.

Fairly easy fix. Nothing to say really. However pay attention, if you use the upper body detector, bounding boxes are not guaranteed to be within the image. The nice thing is though, if you multiply the height with 3, you get a full body bounding box.

[cdondrup]

Nice fix ;) And thank you for taking on the task of merging some of the stuff other people do into our upstream.

[Pandoro]

I like praise, but I have no idea what The second part even means so I certainly didn't do it :x On Feb 25, 2015 3:13 PM, "Christian Dondrup" notifications@github.com wrote:

Nice fix ;) And thank you for taking on the task of merging some of the stuff other people do into our upstream.

— Reply to this email directly or view it on GitHub https://github.com/strands-project/strands_perception_people/pull/137#issuecomment-75965691 .

[cdondrup]

Don't worry. The gratitude was the main message ;)

[lucasb-eyer]

Notice that this makes the detector output bounding-boxes that potentially go outside the image bounds, which did not happen previously. If components building on top of it are not aware of this fact, they will access out-of-memory regions and behave weird and/or randomly crash. Also, this will happen rarely, because if the bounding-box goes outside the image, the template very quickly stops matching.

Just so we are aware of that.

[cdondrup]

This might be interesting for @hawesie and his student project creating 3D bounding boxes. That said, I have no idea if you used the upper body detector in the end. Just to let you know.

[hawesie]

Yes, thanks! I will pass on this information. She has written a node that creates segmented images of people from the tracks, so this will certainly effect her.

[cdondrup]

It should also make it more accurate though and, as @Pandoro said, easier to get the full body bounding box.

[hawesie]

Is this in the latest release?

[cdondrup]

In 0.1.3 which has been released after this was merged, yes.

[lucasb-eyer]

1. In the off-chance your student needs it, here's how to do it.
2. @Pandoro Just above the code you disabled, there's the same story with the bbox width, so I guess we have the same problem with people close to the camera edge. We should quickly test if this happens the next time one of our students runs the upper-body detector.

[Pandoro]

Hmmm, yes that is true, but as the bbox is square, I think it is only determined by the last case, but it is true, we should check this.

On Wed, Feb 25, 2015 at 9:37 PM, Lucas Beyer notifications@github.com wrote:

1. In the off-chance your student needs it, here https://github.com/strands-project/strands_perception_people/blob/indigo-devel/strands_head_orientation/src/main.cpp#L93-L96's how to do it.
2. @Pandoro https://github.com/Pandoro Just above the code you disabled, there's the same story https://github.com/strands-project/strands_perception_people/blob/indigo-devel/upper_body_detector/src/ROI.cpp#L142-L143 with the bbox width, so I guess we have the same problem with people close to the camera edge. We should quickly test if this happens the next time one of our students runs the upper-body detector.

— Reply to this email directly or view it on GitHub https://github.com/strands-project/strands_perception_people/pull/137#issuecomment-76050902 .

[hawesie]

Thanks @lucasb-eyer