This new Object-Recognition Dataset, ‘ObjectNet’, stumped the leading Computer Vision Models

Developed by a team of MIT and IBM researchers, ‘ObjectNet’ is a dataset that tries to solve the complexity of real-world objects. 

Unlike ImageNet, which utilizes pictures taken from Flickr and other online resources, ObjectNet uses pictures taken by freelancers. The ObjectNet has a new kind of vision dataset, borrowing the idea of controls from other areas of science. It doesn’t even have the training set and provides only the test set to expedite the process. The images collected in ObjectNet intentionally shows objects from different viewpoints on new backgrounds. The sample size is 50,000 image test set, same as ImageNet, with advanced features like controls for rotation, background, and viewpoint. It has 313 object classes with 113 overlapping ImageNet.

While some of the leading object detection models were tested on ObjectNet, their accuracy rates fell from a high of 97 percent on ImageNet to just 50-55 percent.


Download dataset:


Dataset Construction

Image Source:

Easy for humans, hard for machines

Image Source: