Cross-dataset Generalization

• virtually no papers demonstrating cross-dataset generalization, e.g. training on ImageNet, while testing on PASCAL VOC
• if our datasets were truly representative of the real world, this would be a very easy thing to do, and would give access to more of the much needed labelled data
• But from our perspective, all the datasets are really trying to represent the same domain – our visual world – and we would like to measure how well or badly they do it.
• Overall the results look rather depressing, as little generalization appears to be happening beyond the given dataset