In my Btech mini project, Sayeem Shaikh, a PhD scholar; Maharshi Vyas, a BTech student and I were working on the Image captioning project using CNNs and RNNs in Torch. We reproduced the Andrej Karpathy results first from his paper. After that, we were tickling with the architecture to work on more data. During the project, the blind data augmentation method developed by us resulted in better BLEU score on 4-gram model than Karpathy et al.'s paper. By blind augmentation, we mean that there was no prior to start with.
Mentor- Manjunath V. Joshi, Professor, DA-IICT, Gandhinagar