Where to put the Image in an Image Caption Generator
This addresses a practical design issue for researchers and engineers building captioning systems, but it is incremental as it compares existing methods without introducing new paradigms.
The paper tackled the problem of where to incorporate image information in an image caption generator, comparing 'injecting' vs. 'merging' image features with RNNs, and found that performance is not significantly harmed by either choice, with the merge architecture allowing up to a fourfold reduction in hidden state size.
When a recurrent neural network language model is used for caption generation, the image information can be fed to the neural network either by directly incorporating it in the RNN -- conditioning the language model by `injecting' image features -- or in a layer following the RNN -- conditioning the language model by `merging' image features. While both options are attested in the literature, there is as yet no systematic comparison between the two. In this paper we empirically show that it is not especially detrimental to performance whether one architecture is used or another. The merge architecture does have practical advantages, as conditioning by merging allows the RNN's hidden state vector to shrink in size by up to four times. Our results suggest that the visual and linguistic modalities for caption generation need not be jointly encoded by the RNN as that yields large, memory-intensive models with few tangible advantages in performance; rather, the multimodal integration should be delayed to a subsequent stage.