Cntk: Object Detection and draw bounding box around object

[mrahmanWM]

Hi,

Let's say the training dataset I have has two classes: traffic sign (traffic sing covered the whole image) and background (no traffic sign). The test dataset contains large images with more than one traffic sign in every image. My goal is to detect those traffic signs from the test images and draw bounding boxes around those. I also want to know the coordinates of the detected traffic sign and measure precision, recall and overlap percentage. I read the Alexnet configuration file and from there I understand what to do to read images. But unfortunately I did not find the configuration to output images with bounding boxes and also the coordinates, precision, recall etc. Any help? Please.

[mrahmanWM]

Hi,

I still could not figure it out? So, any idea about how can I get those things? Only coordinates will work. Please.

[Alexey-Kamenev]

Hello,
We currently does not support such functionality. However, having something like R-CNN (its faster version, to be precise) is in our plans.

[Piso]

Any hopes in seeing faster-rcnn supported any time soon?

[n17s]

Yes it should arrive soon. All our development is public and you can easily search for branches that contain rcnn via the web interface.

[PatrickBue]

Fyi, Fast R-CNN is now part of CNTK:
https://github.com/Microsoft/CNTK/tree/master/Examples/Image/Detection/FastRCNN

[mrahmanWM]

Excellent.

[andreapiso]

Great to have fast-rcnn!

I might be pushing it a bit, but are there plans to get to the _faster_ rcnn as well? :)

[pkranen]

Yes, please stay tuned. Thanks for your interest!

[andreapiso]

Awesome!`

I also found this on the docs:

Other items for future work include: Evaluating alternatives to the currently used selective search implementation. Adding bounding box regression. Implementation of faster R-CNN, i.e. performing ROI generation inside the DNN. Using a more recent DNN topology such as ResNet instead of AlexNet.

This is amazing!

[harish-geeky]

Try Playment.in.