Hi guys,
with the tool camera-capture I’m able to capture single images from the camera live stream.
Because of the parameter list, it seems that it’s possible to select images from a local folder “file://folderofimages” as the input-url. I have already a lot of images and don’t need the camera.
But if I try the “file://” parameter, the tool seems not to work because I got an error from the camera.
Are there any other parameter to set to get the local images working?
Regards,
Stefan
Hi @stefan.hauenstein, if you already have the images or video saved, used a tool like CVAT to annotate them. Then export from CVAT in Pascal VOC format and add a labels.txt file to your dataset. This is assuming you are making an object detection dataset.
If you are doing image classification, you just need to organize your images by subdirectory, where each class has it’s own subdirectory under the train/ val/ test/ sets.
Hi @dusty_nv, man, are you working 24/7 at nvidia :-D ?
Provides the CVAT tool already the folder structure which is used by train_ssd.py?
You are right, I making an object detection database.
Yep it does, export it as Pascal VOC, and you just need to add a labels.txt file with it. The labels.txt should have one class name per line.
Should the labels basically placed in alphabetical order inside the lables.txt file?
For object detection in Pascal VOC format, it does not actually matter the order in labels.txt, but I tend to do it in alphabetical order.
With image classification, alphabetical order is important because the class ID’s are equal to the order of the class directories on disk (which is alphabetical)
Ok, thank you.