While models of fine-grained recognition have made great progress in recent years, little work has focused on a key ingredient of making recognition work: data. We use publicly available, noisy data sources to train generic models which vastly improve upon state-of-the-art on fine-grained benchmarks. First, we present an active learning system using non-expert human raters, and improve upon state-of-the-art performance without any text or other metadata associated with the images. Second, we show that training on publicly-available noisy web image search results achieves even higher accuracies, without using any expert-annotated training data, while scaling to over ten thousand fine-grained categories. We analyze the behavior of our models and data and make a strong case for the importance of data over special-purpose modeling: using only an off-the-shelf CNN, we obtain top-1 accuracies of 92.8\% on CUB-200-2011 Birds, 85.4\% on Birdsnap, 95.9\% on FGVC-Aircraft, and 82.6\% on Stanford Dogs.

利用简单通用的识别方法和网络上的免费嘈杂数据，成功规模扩展到10,000多个类别，不需要手动收集标签，取得超过现有各领域最佳记录的92.3%、85.4%、93.4%和80.8% 的top-1准确率

噪声数据对精细识别的非合理有效性