NoW

홈페이지 · 논문 2편

Noise of Web (NoW) is a challenging noisy correspondence learning (NCL) benchmark for robust image-text matching/retrieval models. It contains 100K image-text pairs consisting of website pages and multilingual website meta-descriptions (98,000 pairs for training, 1,000 for validation, and 1,000 for testing). NoW has two main characteristics: *without human annotations and the noisy pairs are naturally captured*. The source image data of NoW is obtained by taking screenshots when accessing web pages on mobile user interface (MUI) with 720 $\times$ 1280 resolution, and we parse the meta-description field in the HTML source code as the captions. In NCR (predecessor of NCL), each image in all datasets were preprocessed using Faster-RCNN detector provided by Bottom-up Attention Model to generate 36 region proposals, and each proposal was encoded as a 2048-dimensional feature. Thus, following NCR, we release our the features instead of raw images for fair comparison. However, we can not just use detection methods like Faster-RCNN to extract image features since it is trained on real-world animals and objects on MS-COCO. To tackle this, we adapt APT as the detection model since it is trained on MUI data. Then, we capture the 768-dimensional features of top 36 objects for one image. Due to the automated and non-human curated data collection process, the noise in NoW is highly authentic and intrinsic. The estimated noise ratio of this dataset is nearly 70%.