Hi @savunmasad, I believe those are the two that I packaged for Hello AI World - those are subsets of larger datasets (e.g. PlantCLEF). The Object Detection Training portion of the tutorial also has a tool that downloads selected categories from the Open Images dataset.
If you are looking for more classification data, those are pretty easy to find or create, because you just need to organize the images in a simple directory structure like the one shown here:
And can i add a new object to cocodataset
Do you mean adding a new class to MS COCO for object detection? In theory, it is possible - you would need to add the object in MS COCO format. The tools I created in jetson-inference work in PASCAL VOC format, and the PyTorch SSD training script work with PASCAL VOC format and Open Images format (but not COCO).
If you are just using COCO for image recognition as opposed to detection, that should be easier to handle.
Yes new class.
How can i do that
The directory structure for image recognition/classification datasets in PyTorch is outlined here:
So the MS COCO images would need to be organized like that (probably by some type of script), and then you could easily add your own class by adding an additional class folders.
One issue that occurred to me, is that the COCO dataset frequently has multiple different objects per image (whereas classification is typically one). COCO is detection, segmentation, and image captioning.