Research map / Vision-language and multimodal / Human and embodied modeling
Partially Non-Autoregressive Image Captioning
Zhengcong Fei
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 |
|---|---|---|---|
| Increase the number of parallel word groups from K = 2 to K = 5 | PNAIC · MS COCO test | SpeedUp: 2.17× -> 3.59× | Improved |
| Increase the number of parallel word groups from K = 5 to K = 10 | PNAIC · MS COCO test | CIDEr: 128.5 -> 125.2 | Hurt |
| Increase the number of parallel word groups from K = 5 to K = 10 | PNAIC · MS COCO test | SpeedUp: 3.59× -> 5.43× | Improved |
| Use curriculum-based group-length prediction rather than fixed f(m) = 1 (ablation) | PNAIC · MS COCO validation (K = 5) | CIDEr: 124.3 -> 128.5 | Improved |
| Use curriculum-based group-length prediction rather than fixed f(m) = 0 (ablation) | PNAIC · MS COCO validation (K = 5) | CIDEr: 128.3 -> 128.5 | No clear effect |
| Use curriculum-based group-length prediction rather than a linear f(m) schedule (ablation) | PNAIC · MS COCO validation (K = 5) | CIDEr: 128.2 -> 128.5 | No clear effect |
| Add pseudo repetitive groups for invalid-group deletion training (p = 0.5) (ablation) | PNAIC · MS COCO validation (K = 5) | CIDEr: 122.5 -> 128.5 | Improved |
| Increase invalid-group injection probability from p = 0.5 to p = 0.9 (ablation) | PNAIC · MS COCO validation (K = 5) | CIDEr: 128.5 -> 125.5 | Hurt |
| Set invalid-group injection probability to p = 1.0 rather than p = 0.5 (ablation) | PNAIC · MS COCO validation (K = 5) | CIDEr: 128.5 -> 124.8 | Hurt |
| Increase invalid-group injection probability from p = 0 to p = 0.1 (ablation) | PNAIC · MS COCO validation (K = 5) | CIDEr: 122.5 -> 126.8 | Improved |
| Increase invalid-group injection probability from p = 0.1 to p = 0.3 (ablation) | PNAIC · MS COCO validation (K = 5) | CIDEr: 126.8 -> 127.6 | Improved |
| Increase invalid-group injection probability from p = 0.5 to p = 0.7 (ablation) | PNAIC · MS COCO validation (K = 5) | CIDEr: 128.5 -> 127.9 | Hurt |
Closest papers by meaning
- Length-Controllable Image Captioning (ECCV 2020 · similarity 0.88)
- Non-Autoregressive Coarse-to-Fine Video Captioning (AAAI 2021 · similarity 0.87)
- Image Captioning With Controllable and Adaptive Length Levels (TPAMI 2023 · similarity 0.85)
- FastHybrid: Accelerating Hybrid Autoregressive Image Generation with Lookahead and Guided Decoding (CVPR 2026 · similarity 0.85)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.84)
- Image Captioning with Multi-Context Synthetic Data (AAAI 2024 · similarity 0.84)
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.
