ObjectNet (720P Shortest Edge, ImageNet-1k Overlap)

A webp (lossless) encoded version of ObjectNet-1.0 resized to shortest edge = 720 pixels. Containing only the 113 classes that overlap with ImageNet-1k.

License / Usage Terms

ObjectNet is free to use for both research and commercial applications. The authors own the source images and allow their use under a license derived from Creative Commons Attribution 4.0 with only two additional clauses.

ObjectNet may never be used to tune the parameters of any model.
Any individual images from ObjectNet may only be posted to the web including their 1 pixel red border. If you are using ObjectNet, please cite our work, the citation appears at the bottom of this page. Any derivative of ObjectNet must contain attribution as well.

About

What is ObjectNet?

A new kind of vision dataset borrowing the idea of controls from other areas of science.
No training set, only a test set! Put your vision system through its paces.
Collected to intentionally show objects from new viewpoints on new backgrounds.
50,000 image test set, same as ImageNet, with controls for rotation, background, and viewpoint.
313 object classes with 113 overlapping ImageNet
Large performance drop, what you can expect from vision systems in the real world!
Robust to fine-tuning and a very difficult transfer learning problem

Why the Red Borders / How do I recognize if an image is in ObjectNet?

As training sets become huge, the risk that test and training sets overlap is serious. We provide ObjectNet with a 2 pixel red border around each image which must be removed before performing inference. The ObjectNet license requires that if you post images from ObjectNet to the web, you include this border. Any time you see an image with a solid 2 pixel red border, that's an indication it's in someone's test set and you should be careful about training on it. Reverse image search will allow you to figure out which test set it is from.

NOTE: original ObjectNet PNG files actually have a 2 pixel red border while their descriptions say 1.

Preprocessing Steps for This timm Version

Remove 2-pixel red border.
Resize images to shortest edge = 720 pixels (if shortest edge larger than this).
Add back 1-pixel red border.
Re-encode PNG images with lossless WebP. In total, ~76% of original size.
Add imagenet_labels and imagenet_synsets consisting of lists of ImageNet-1k classes that overlap with ObjectNet class.
Filter out classes that do not overlap with ImageNet-1k.

Citation

@incollection{NIPS2019_9142,
  title = {ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models},
  author = {Barbu, Andrei and Mayo, David and Alverio, Julian and Luo, William and Wang, Christopher and Gutfreund, Dan and Tenenbaum, Josh and Katz, Boris},
  booktitle = {Advances in Neural Information Processing Systems 32},
  editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
  pages = {9448--9458},
  year = {2019},
  publisher = {Curran Associates, Inc.},
  url = {http://papers.nips.cc/paper/9142-objectnet-a-large-scale-bias-controlled-dataset-for-pushing-the-limits-of-object-recognition-models.pdf}
}

Datasets:

timm
/

objectnet-720p-in1k