Document localization is a primary step in intelligent document analysis. The document images captured by smartphones in nature scene inevitably contain multiple documents, but previous methods rarely focus on multi-document localization. In this paper, we propose a bottom-up multi-document localization model with a multi-document dataset. Specifically, we design several high-to-low resolution branches and repeat fusion across parallel multi-resolution representations to localize corners more accurately. We also introduce an enhanced feature module to improve feature representation around small or blurry document corners. Then, we adopt an associative embedding-based grouping mechanism which generates unique embedding vectors for each corner and clusters the corners automatically to generate complete document instances more efficiently compared with other top-down approaches. For evaluation, we collect a multi-document dataset, containing 24,738 document images in various unstructured environments. Our experiments demonstrate that the proposed baseline model trained on this dataset achieves high accuracy and efficiency in detecting both single and multiple documents. The code and dataset are available on GitHub at https://github.com/TanQiuman/Bottom-Up-Multi-document-Localization .