A single-object version of the ImageNet2012 dataset

The dataset includes 544,546 images in the training dataset and 50,000 images in the validation dataset. We removed the background of each image by setting pixels outside the bounding box to 255 (i.e., white color).

For images containing multiple bounding boxes, we randomly selected one bounding box as our target. Note that the retinal size of objects remained unchanged with only the background was removed from the original images. All images belong to 1,000 categories, which are the same as the original ImageNet2012 dataset.