ObjectNet is a large-scale image test set collected through crowdsourcing. Since the images in ObjectNet are all from the real world, the target objects have different shapes, and the background, rotation angle and imaging perspective are all random, which may seriously reduce the recognition effect.

The test set includes 50,000 images with different rotations, backgrounds, and perspectives, covering 313 object categories, 113 of which are consistent with the ImageNet test set.