Dataset

The WebVision dataset is designed to facilitate the research on learning visual representation from noisy web data. Our goal is to disentangle the deep learning techniques from huge human labor on annotating large-scale vision dataset. Along this goal, researchers consider different strategies by using unlabeled data, semi-supervised data, or loosely labeled web data. Specifically, web images/videos are naturally accompanied with abundant meta information, which could be used as weak supervision for learning deep neural networks.

To investigate the capacity of web data for learning robust visual representation, we construct the large scale  WebVision dataset by crawling images from Internet. The same 1,000 concepts as the ILSVRC 2012 dataset are used for querying images, such that a bunch of existing approaches can be  directly investigated and compared to the models trained from the ILSVRC 2012 dataset. In total, we obtain about 3.5 million images. The textual information accompanied with those images (e.g., caption, user tags, or description) are also provided as additional meta information. The preliminary results of a simple baseline show that this dataset is able to learn robust representation which achieves comparable performance as the model learnt from the human annotated ILSVRC 2012 dataset on several vision tasks (see details below). We provide this dataset for academic research and education purposes.

Click here for a detailed description and download instructions

Challenge

Observing the success of the PASCAL-VOC, ILSVRC, COCO and Places competitions to the advance of computer vision community, we propose to organize a competition on learning from web images to promote the development of this field. We divide the WebVision dataset into training, validation and test sets. The validation and test sets are cleaned by human, and are only used for evaluation purpose in the challenge.

Two tracks of tasks will be organized with this collected training set: 1) image classification on annotated test data from 1000 categories: training is only allowed on this collected training set, and 50 human annotated test image per category are labeled for benchmarking. 2) transferred learning on standard benchmark dataset (PASCAL): the collected training set and PASCAL training set is available for learning and adapting. Standard PASCAL testing set and protocol will be used for benchmarking.