Reading bounding box shapes as input to 'label' layer

[Manohar]

How to create a database to store the data of bounding boxes in case of object detection ?

Say for eg:

Image_1.png  has 1 object with co-ordinates [xmin1, ymin1, xmax1, ymax1, C1]  -- 'C1' is the label of the class to which the object belongs to. 

Image_2.png  has 2 objects with co-ordinates [xmin1, ymin1, xmax1, ymax1, C1], [xmin2, ymin2, xmax2, ymax2 ,C2]

Given that an image can have any number of objects each object being specified by 5 values, how do I create a database for training an object detection model ?

[Romain DAMBREVILLE]

Hi,
I don't know what network you are using, but in my case (faster-rcnn) I followed this tuto https://huangying-zhan.github.io/2016/09/22/detection-faster-rcnn.html#Training%20on%20new%20dataset, and took the dataset as an example.
Look at the Annotation folder in root_folder/data/VOCdevkit/VOC2007/, you will find all the xml files containing annotations (labels).
Then I reproduced the same format for my data and managed to feed it by editing some of the python files (cf tuto).
Hope it can help you.
Romain

[Manohar]

I'm using SqueezeDet architecture and the dataset is PASCAL VOC

[Manohar]

And the model uses YOLO loss function

On Monday, June 19, 2017 at 10:02:23 AM UTC+5:30, Manohar wrote: