|
Object Recognition_Template folder samples Frederick Prete from United States [6 posts] |
7 year
|
Colleagues:
By any chance are there sample folders of template images available on-line for down load that can be used in experimenting with the Object Recognition module?
Thanks in advance,
Dr. P.
|
|
|
Steven Gentner from United States [1446 posts] |
7 year
|
Dr. P,
Not that we can make available. Typically google image search is a good option for finding and creating such a database depending on what objects you are interested in detecting. The issue with using Google (or other search engine) is that the images are typically not part of the public domain and thus should only be used for private use.
Alternatively, there are many data sets produced from some universities/organizations specific to certain competitions. For example, if you check ImageNet
http://image-net.org/challenges/LSVRC/2016/
you can find an annotated data set there.
http://imagenet.stanford.edu/synset?wnid=n02892767
STeven.
|
|
|
Frederick Prete from United States [6 posts] |
7 year
|
Thank you, STeven.
So, in creating a folder of test images, I assume that the images are simply saved in a folder as jpegs of some target object against both plain and heterogeneous backgrounds. (As per the links that you provided.) RoboRealm is then trained to that folder and the identified item is labeled as per the folder name.
Am I missing something, or is it that straightforward?
FRP
|
|
|
Steven Gentner from United States [1446 posts] |
7 year
|
FRP,
Yes, that's basically it. The one change is that the images should be only of the target to detect and NOT include any background whatsoever. If you press the Add Object button in the module it will allow you to select a specific area and then just save that area as an image within a specified folder. So all "templates" are just regular jpg images that you can move/edit/cut/paste/etc. using the file system.
The reason we don't want any background is that this is NOT a NN or GP type system that is learning the object. Its more like SURF or SIFT techniques in that a single image is used to attempt to identify a particular planar object.
STeven.
|
|