WebVision Database: Visual Learning and Understanding from Web Data

• DOI: 10.48550/arxiv.1708.02862
• Publication Date: 2017-01-01
• Authors: Wen Li; Limin Wang; Li Wei; Eirikur Agustsson; Luc Van Gool
• Keywords: Benchmark (surveying); Domain Adaptation; Data set

In this paper, we present a study on learning visual recognition models from large scale noisy web data. We build new database called WebVision, which contains more than $2.4$ million images crawled the Internet by using queries generated 1,000 semantic concepts of benchmark ILSVRC 2012 dataset. Meta information along with those (e.g., title, description, tags, etc.) are also crawled. A validation set and test containing human annotated provided to facilitate algorithmic development. Based our database, obtain few interesting observations: 1) sufficient for training good deep CNN model recognition; 2) learnt WebVision exhibits comparable or even better generalization ability one trained dataset when being transferred datasets tasks; 3) domain adaptation issue (a.k.a., bias) is observed, means can be used as largest adaptation. Our relevant studies in work would benefit advance state-of-the-art minimum supervision based