As image datasets become ubiquitous, the problem of ad-hoc searches over image data is increasingly important. In particular, many tasks, such as constructing datasets for training and testing object detectors, require finding ad-hoc objects or scenes within large image datasets. Existing approaches for searching image datasets rely on rigid categories or assume fully accurate models trained on the data are available ahead of time. In contrast, SeeSaw is a system for interactive ad-hoc searches in image datasets that does not assume a pre-defined set of categories in advance. SeeSaw users can start a search using text, and then provide feedback to SeeSaw in the form of box annotations on previous results. Using this input, SeeSaw refines the results it returns to help users locate images of interest in their data. Behind the scenes, SeeSaw uses several optimizations to transform the user's feedback into better results. We evaluate SeeSaw against a state of the art baseline that does not take advantage of user feedback and find SeeSaw can increase search quality metrics up 4x for the harder search queries, and significantly on almost all queries, even those where the baseline performs well.