Research map / Detection and segmentation / CNN features
GRIT: Faster and Better Image Captioning Transformer Using Dual Visual Features
Van-Quang Nguyen, Masanori Suganuma, Takayuki Okatani
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 end-to-end training of backbone and detector with the caption generator (ablation) | GRIT · COCO Karpathy test | CIDEr: 139.6 -> 144.2 | Improved |
| pretrain detector on four datasets instead of Visual Genome only (ablation) | GRIT · COCO Karpathy test | CIDEr: 142.3 -> 144.2 | Improved |
| add region features from detector pretraining on top of ImageNet grid-only backbone (ablation) | GRIT · COCO Karpathy test | CIDEr: 135.5 -> 142.3 | Improved |
| increase number of region queries from 50 to 150 (ablation) | GRIT (VG-trained) · COCO Karpathy test | CIDEr: 141.4 -> 142.3 | Improved |
| increase number of region queries from 100 to 150 (ablation) | GRIT (VG-trained) · COCO Karpathy test | BLEU-4: 41.5 -> 41.9 | No clear effect |
| add region features to grid-only concatenated cross-attention (ablation) | GRIT concatenated attention · COCO Karpathy test | CIDEr: 142.1 -> 143.1 | Improved |
| use region features alone instead of grid alone (ablation) | GRIT concatenated attention · COCO Karpathy test | CIDEr: 142.1 -> 142.9 | Improved |
| use parallel cross-attention instead of concatenation (ablation) | GRIT · COCO Karpathy test | CIDEr: 143.1 -> 144.2 | Improved |
| use parallel cross-attention instead of sequential (grid then region) (ablation) | GRIT · COCO Karpathy test | CIDEr: 144.0 -> 144.2 | Improved |
| order sequential attention grid first then region rather than region first (ablation) | GRIT sequential attention · COCO Karpathy test | CIDEr: 143.6 -> 144.0 | Improved |
| add sigmoid gating to parallel cross-attention (vs identity) (ablation) | GRIT parallel attention · COCO Karpathy test | CIDEr: 143.9 -> 144.2 | Improved |
| add sigmoid gating to parallel cross-attention (vs identity) (ablation) | GRIT parallel attention · COCO Karpathy test | BLEU-4: 41.6 -> 42.4 | Improved |
Later papers that compared with this method
4 later accepted papers report +GRIT (20M) as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- Synthetic Object Compositions for Scalable and Accurate Learning in Detection, Segmentation, and Grounding (CVPR 2026 · better on 3 of 3 shared results)
- DetCLIPv3: Towards Versatile Generative Open-vocabulary Object Detection (CVPR 2024 · better on 1 of 1 shared result)
- Segment and Caption Anything (CVPR 2024 · better on 1 of 1 shared result)
- NExT-Chat: An LMM for Chat, Detection and Segmentation (ICML 2024 · better on 0 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GRiT: A Generative Region-to-text Transformer for Object Understanding (ECCV 2024 · similarity 0.89)
- End-to-End Transformer Based Model for Image Captioning (AAAI 2022 · similarity 0.88)
- Dual-level Collaborative Transformer for Image Captioning (AAAI 2021 · similarity 0.86)
- From Show to Tell: A Survey on Deep Learning-Based Image Captioning (TPAMI 2022 · similarity 0.86)
- Convolutional Image Captioning (CVPR 2018 · similarity 0.86)
- Gradient-based Visual Explanation for Transformer-based CLIP (ICML 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.
