Research map / Video / Video-language understanding
Hierarchical LSTMs with Adaptive Attention for Visual Captioning
Lianli Gao, Xiangpeng Li, Jingkuan Song, Heng Tao Shen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 temporal attention to the basic LSTM (SA vs basic LSTM) (ablation) | basic LSTM · MSVD (ResNet-152) | CIDEr: 69.9 -> 72.0 | Improved |
| add adaptive attention to the basic LSTM (ablation) | basic LSTM · MSVD (ResNet-152) | CIDEr: 69.9 -> 71.5 | Improved |
| add adaptive attention to hLSTMt (ablation) | hLSTMt · MSVD (ResNet-152) | BLEU@4: 52.1 -> 53.0 | Improved |
| add adaptive attention to hLSTMt (ablation) | hLSTMt · MSVD (ResNet-152) | CIDEr: 73.5 -> 73.8 | Improved |
| use hierarchical (two-layer) LSTMs instead of a single LSTM with temporal attention (ablation) | SA · MSVD (ResNet-152) | CIDEr: 72.0 -> 73.5 | Improved |
| use ResNet-152 features instead of C3D features (ablation) | hLSTMat · MSVD (encoder study) | BLEU@4: 47.5 -> 53.0 | Improved |
| use deeper ResNet encoder (ResNet-50 to ResNet-101 to ResNet-152) (ablation) | hLSTMat · MSVD (encoder study) | BLEU@4: 49.0 -> 50.8 -> 53.0 | Improved |
| use Two-stream fusion instead of concatenation fusion of C3D and ResNet-152 (ablation) | hLSTMat · MSVD (C3D+ResNet-152 fusion) | CIDEr: 65.1 -> 71.3 | Improved |
| use ParA parallel adaptive attention instead of concatenation fusion (ablation) | hLSTMat · MSVD (C3D+ResNet-152 fusion) | BLEU@4: 52.5 -> 53.6 | Improved |
| add C3D motion features to ResNet-152 appearance features (Two-stream) (ablation) | hLSTMat · MSVD (C3D+ResNet-152 fusion) | CIDEr: 73.8 -> 71.3 | Mixed |
| replace C3D with Inception-v3 in the two-stream model (ablation) | Two-stream · MSVD | BLEU@4: 54.0 -> 56.2 | Improved |
| use spatial region attention instead of temporal attention (ablation) | hLSTMat · MSVD (spatial attention) | CIDEr: 73.8 -> 72.3 | Mixed |
Closest papers by meaning
- Hierarchical Attention Network for Image Captioning (AAAI 2019 · similarity 0.89)
- Knowing When to Look: Adaptive Attention via a Visual Sentinel for Image Captioning (CVPR 2017 · similarity 0.88)
- Adaptively Aligned Image Captioning via Adaptive Attention Time (NeurIPS 2019 · similarity 0.88)
- RSTNet: Captioning With Adaptive Attention on Visual and Non-Visual Words (CVPR 2021 · similarity 0.88)
- MAMS: Model-Agnostic Module Selection Framework for Video Captioning (AAAI 2025 · similarity 0.86)
- Visual Attention Never Fades: Selective Progressive Attention ReCalibration for Detailed Image Captioning in Multimodal Large Language Models (ICML 2025 · similarity 0.86)
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.
