Your dataset does not need to contain images from official databases providing these types, like ImageNet or Pascal VOC, but it needs to adhere to the supported dataset formats.
To learn how to download and prepare datasets, refer to Download and Cut Datasets.
NOTE: For COCO datasets, only the object-detection and instance-segmentation tasks are supported.
ImageNet is a well-known dataset used to train classification models. It consists of an annotation file and images:
The annotation file is organized as follows:
Pascal VOC is a well-known dataset used to train object-detection and semantic-segmentation models. Pascal VOC datasets consist of several folders containing annotation files and image indices.
A Pascal VOC dataset archive is organized as follows:
COCO dataset is used for object detection, instance segmentation, person keypoints detection, stuff segmentation, and caption generation.
A COCO dataset is downloaded as two separate archives, but you have to create one archive based on them as described in the Download COCO Dataset section in Download and Cut Datasets. To upload a COCO dataset to the DL Workbench, make sure the archive contains the following files:
The JSON file with annotations is organized as follows:
Common Semantic Segmentation is a custom dataset type for semantic segmentation models. The type is aimed to simplify the structure provided by Pascal VOC.
A Common Semantic Segmentation dataset archive consists of folders with images and masks, and a JSON file with metainformation:
The JSON metainformation file is organized as follows:
Common Super-Resolution is a custom dataset type for super-resolution, image-inpainting, and style-transfer models.
The archive consists of three separate folders for high-resolution images, low-resolution images, and upsampled low-resolution images:
Unannotated datasets do not contain annotations, so models in the configurations that use unannotated datasets can be calibrated only in the simplified mode of the Default Calibration method and cannot be used for accuracy measurement.
The archive is organized as follows: