Darkflow support for the VOC standard

[4F2E4A2E]

Hello there @thtrieu!

Thank you very much for this amazing repo. It has been a lot of fun working with darknet and darkflow and you made it all happen! 👍

I am sure that, only in repo and issues management a lot of hours have flown, so thank you sir in many ways, i will try to keep it short:

What i want to learn and would love to undestand is how VOC is really being integrated or regarded into darknet &/ darkflow.

VOC-Standard / VOC-Dataset folder structure:

-Annotations
-ImageSets
--Action
--Layout
--Main
--Segmentation
-JPEGImages
-SegmentationClass
-SegmentationObject

To my best knowledge: The current source code [1] in darkflow &/ darknet does not contain any references or is in any way processing the VOC-Standard in something else other than

• -Annotations (class / label annotation) and
• -JPEGImages (images)

If this is correct, than a enormous poetential in the training &/ recognition is ... lost? Not being used? Processed? Not to mention the annotation files itself, which contains flags like:

• truncated
• occluded
• difficult, etc

Are they those information being processed? If yes how!? There seams to be nothing on this in the source code.

I am hoping that at least SegmentationClasses and SegmentationObjects are being somehow involved and being processed by the model alias yolo-voc.weight files.

I kindly ask you to totally dismantle this guess or to enlighten us. That would be really great and helpfull!

This way:

• darkflow would still rock
• contributors could extend the functionality in order to fully support the VOC-Standard
• knowing the limitations, datasets would be prepared much faster and slicker regarding the VOC-features supported by darkflow only

1:

• https://github.com/pjreddie/darknet/search?utf8=%E2%9C%93&q=truncated&type=
• https://github.com/thtrieu/darkflow/blob/master/darkflow/defaults.py
• https://github.com/thtrieu/darkflow/search?utf8=%E2%9C%93&q=truncated&type=

Thank you in advance!

[4F2E4A2E]

Any hint?

[rlan]

TLDR I think darkflow only supports object detection eg no segmentation AFAIK.

Here is how I understood the history of darkflow. It is a port of darknet which is the code base for the paper by Redmon et al, YOLO. The paper topic is on object detection. On Sun, Oct 8, 2017 at 6:36 AM 4F2E4A2E notifications@github.com wrote:

Any hint?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/thtrieu/darkflow/issues/406#issuecomment-334967266, or mute the thread https://github.com/notifications/unsubscribe-auth/ABUpdFzbinUM0PQh4OWfay9-xjX8zGmUks5sp-7NgaJpZM4Pg8Ai .

[4F2E4A2E]

It seams that darknet does make use of the segmentated images containing the ground truth. Maybe this is only not yet implemented &/ configured on darkflow?