objectnet-in1k

# ObjectNet (ImageNet-1k Overlapping)

A webp (lossless) encoded version of [ObjectNet-1.0](https://objectnet.dev/index.html) at original resolution, containing only the images for the 113 classes that overlap with ImageNet-1k classes.

## 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.

1. **ObjectNet may never be used to tune the parameters of any model.**
2. **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 ObjectNet

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
1. Re-encode PNG images with lossless WebP (~32% reduction in size), keeping red border.
2. Add `imagenet_labels` and `imagenet_synsets` consisting of lists of ImageNet-1k classes that overlap with ObjectNet class.
3. Remove all ObjectNet image classes without ImageNet-1k labels.

## Citation
```bibtex
@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}
}
```

# ObjectNet (ImageNet-1k 重叠子集)

这是[ObjectNet-1.0](https://objectnet.dev/index.html)的无损WebP编码版本，保留原始分辨率，仅包含与ImageNet-1k类别重叠的113个类别的图像。

## 许可与使用条款

ObjectNet可免费用于研究与商业用途。原作者保留源图像的著作权，其使用许可基于知识共享署名4.0协议（Creative Commons Attribution 4.0），仅附加两条额外条款：
1. **绝对禁止使用ObjectNet微调任何模型的参数**。
2. **仅可将带有1像素红色边框的ObjectNet单张图像发布至网络**。
若使用ObjectNet，请引用本文献，引用信息见本页面底部。任何基于ObjectNet的衍生作品也必须保留原作者署名。

## 关于ObjectNet

什么是ObjectNet？
* 一款借鉴其他科学领域控制变量思想的新型视觉数据集。
* 仅包含测试集，无训练集！可用于全面测试你的视觉系统性能。
* 采集初衷是有意呈现处于全新视角与背景下的目标物体。
* 包含50000张图像的测试集，与ImageNet规模一致，针对旋转、背景与视角设置了控制变量。
* 共313个目标类别，其中113个与ImageNet-1k重叠。
* 模型在该数据集上性能会出现显著下降，这正是视觉系统在真实场景中会遇到的表现！
* 对微调鲁棒性差，是极具挑战性的迁移学习任务。

## 为何设置红色边框 / 如何识别图像属于ObjectNet？

随着训练集规模不断扩大，测试集与训练集出现重叠的风险日益严峻。我们为每张ObjectNet图像添加了2像素宽的红色边框，推理前必须移除该边框。ObjectNet许可协议要求，若将ObjectNet图像发布至网络，必须保留该边框。任何带有纯色2像素红色边框的图像，均属于某测试集的成员，因此需注意切勿在训练中使用此类图像。通过反向图像搜索即可确定其所属的测试集。

> 注意：原始ObjectNet的PNG文件实际带有2像素红色边框，但其文档描述为1像素。

## 本timm版本的预处理步骤

1. 使用无损WebP格式重新编码PNG图像（文件体积减少约32%），保留红色边框。
2. 添加`imagenet_labels`与`imagenet_synsets`列表，分别存储与ObjectNet类别重叠的ImageNet-1k类别的标签与同义词集。
3. 移除所有无对应ImageNet-1k标签的ObjectNet图像类别。

## 引用

bibtex
@incollection{NIPS2019_9142,
title = {ObjectNet: 用于突破目标识别模型极限的大规模偏差控制数据集},
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'Alché-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}
}