Research map / Vision-language and multimodal / Human and embodied modeling
Review Networks for Caption Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add review network (tied weights) on top of attentive encoder-decoder (ablation) | attentive encoder-decoder with VGGNet · MSCOCO offline | BLEU-4 / METEOR / CIDEr: BLEU-4 0.278->0.282; METEOR 0.229->0.233; CIDEr 0.840->0.852 | Helped |
| add discriminative supervision (word occurrence multi-label margin loss) to review network (ablation) | Review Net · MSCOCO offline | CIDEr: CIDEr 0.852->0.879; BLEU-4 0.282->0.287 | Helped |
| add discriminative supervision to attentive encoder-decoder (ablation) | attentive encoder-decoder · MSCOCO offline, no beam search | CIDEr: CIDEr 0.793->0.811 | Helped |
| untie reviewer unit weights (vs. RNN-style shared weights) (ablation) | Review Net + Disc Sup · MSCOCO offline | BLEU-4 / CIDEr: BLEU-4 0.287->0.290; CIDEr 0.879->0.886; METEOR flat (0.238->0.237) | Helped |
| reduce review steps to T_r=4 or 16 instead of 8 (ablation) | Review Net · MSCOCO development set | BLEU-4 / CIDEr: about -0.003 BLEU-4 and -0.01 CIDEr vs T_r=8 | Hurt |
| add review network with RNN encoder and attentive output reviewer over attentive encoder-decoder baselines | source code captioning (HabeasCorpus) · HabeasCorpus | LLH and CS-1 to CS-5: LLH -5.14->-5.06; CS-1 0.2716->0.2889; CS-5 0.3651->0.3840 | Helped |
| use review network in place of attentive encoder-decoder on MSCOCO test server (ensemble, Inception-v3) | Inception-v3 encoder · MSCOCO test server | CIDEr: CIDEr 0.969 vs 0.958 for Semantic Attention (best prior); BLEU-4 slightly below (0.597 vs 0.599) | Helped |
Closest papers by meaning
- Deliberate Attention Networks for Image Captioning (AAAI 2019 · similarity 0.89)
- Regularizing RNNs for Caption Generation by Reconstructing the Past With the Present (CVPR 2018 · similarity 0.88)
- Improving Image Captioning with Conditional Generative Adversarial Nets (AAAI 2019 · similarity 0.86)
- Temporal Deformable Convolutional Encoder-Decoder Networks for Video Captioning (AAAI 2019 · similarity 0.86)
- Reconstruction Network for Video Captioning (CVPR 2018 · similarity 0.85)
- Learning to Guide Decoding for Image Captioning (AAAI 2018 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.