SCAN

Why we included this project

If you work on cross-modal retrieval or image captioning, this is the reference implementation of the Stacked Cross Attention Network, the ECCV 2018 paper that introduced a way to align images with natural-language descriptions. The repo includes training and evaluation scripts, vocabulary tooling, and the exact hyperparameters used to reproduce the paper's results on MS-COCO and Flickr30K, so it's a solid base for experimenting with attention-based image-text matching without rebuilding the architecture. It builds on bottom-up attention features and a bidirectional GRU text encoder, which makes it a useful template for building or benchmarking your own retrieval models. Since it's research code tied to a specific paper, treat it as a reference baseline and learning resource rather than a production service.