Reproduce “Image Captioning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Vikram Mullachery, Vishal Motwani

Mullachery and Motwani investigated architectural modifications for CNN-RNN image captioning models, demonstrating improved BLEU_4 and CIDEr scores over a Neuraltalk2 baseline by increasing LSTM decoder depth and incorporating a ResNet encoder. The work also identified challenges in applying static image captioning to dynamic video frames.

Key insights

  • Increasing the depth of the LSTM decoder to two layers (Model 3) provided the most significant performance improvement in image captioning, outperforming both a single-layer baseline and a deeper four-layer configuration.
  • Replacing the VGGNet encoder with a ResNet (Model 5) yielded better image feature encoding and improved captioning scores, indicating the benefit of more advanced CNN architectures.
  • Directly applying image captioning models to video frames leads to temporal inconsistencies, such as abrupt caption changes and sensitivity to minor visual shifts, highlighting the necessity for temporal modeling in video description.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.