An automatically constucted dataset for augmenting ILSVRC-2012 dataset.
Details
AutoDA is constructed to augment existing dataset ILSVRC-2012. The dataset is labeled by both Web and DCNN. Web provides massive images with rich contextual information, while DCNN well-trained on ILSVRC-2012 are used to label these images and filter out noisy images. Meanwhile, the rich contextual information from Web ensures DCNN to achieve high labeling accuracy with relatively low confidence threshold. Together, we can augment the ILSVRC-2012 dataset in a scalable, accurate, and informative way.
The AutoDA dataset contains 12.5 million images crawled from Web without URL or domain restriction. Since AutoDA is built to augment ILSVRC-2012, both datasets share the same 1000 categories. The number of images per category of AutoDA and ILSVRC-2012 is shown in Fig 1. In addition, AutoDA also provides texutal information (relevant textual descriptions accompanied with image in webpage) for each image. With the new dataset, we wish it will help push forward the research in Object Recognition, Image Caption and Visual Question and Answer, also inspire new research directions such as automatic data labeling and dataset compression.
Dataset Downloads
Before you download the AutoDA dataset, please agree the following terms of access:
You have requested permission to use the AutoDA database. In exchange for such permission, you hereby agrees to the following terms and conditions: 1. The database can only be used for non-commercial research and educational purposes. 2. The authors of the database make no representations or warranties regarding the Database, including but not limited to warranties of non-infringement or fitness for a particular purpose. 3. You accepts full responsibility for your use of the Database and shall defend and indemnify the Authors of AutoDA, against any and all claims arising from your use of the Database, including but not limited to your use of any copies of copyrighted images that you may create from the Database. 4. You may provide research associates and colleagues with access to the Database provided that they first agree to be bound by these terms and conditions. 5. The authors of AutoDA reserve the right to terminate your access to the Database at any time. 6. If you are employed by a for-profit, commercial entity, your employer shall also be bound by these terms and conditions, and you hereby represents that you are authorized to enter into this agreement on behalf of such employer. 7. The Intellectual Property Right Law of the People's Republic of China apply to all disputes under this agreement.
Models
Below table shows the performance and download link of the models we trained on AutoDA.
| DCNN | ILSVRC-2012 | AutoDA | Merged | Merged (No dropout) |
| AlexNet | 60.36 (82.38) | 56.58 (78.57) | 62.71 (83.71) [Model] | 61.72 (82.62) [Model] |
| ResNet-50 | 74.44 (92.11) | 70.17 (88.09) | 77.36 (93.29) [Model] |
The tools for feature extraction is avaiable [Here]