Which computer vision datasets are often used in semi-supervised learning and/or transductive learning?

Those are both fully labeled datasets so they're usually not used for semi-supervised learning. Adam Coates has done some semi-supervised learning experiments on CIFAR-10 by throwing out most of the labels. See his AISTATS 11, ICML 11, and NIPS 11 paper. He also uses the STL-10 dataset that he created but so far as I know, that hasn't caught on with other labs yet. You might look into what Ruslan Salakhutdinov used for his recent one-shot-learning work.