Research map / Detection and segmentation / Open-set and few-shot detection
VinVL: Revisiting Visual Representations in Vision-Language Models
Check 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 |
|---|---|---|---|
| Replace Anderson's visual features with VinVL features | Oscar / Oscar+ vision-language models · VQA test-dev | accuracy: 73.16 -> 75.95 | Helped |
| Replace Anderson's visual features with VinVL features | Oscar / Oscar+ vision-language models · GQA test-dev | accuracy: 61.58 -> 65.05 | Helped |
| Replace Anderson's visual features with VinVL features | Oscar / Oscar+ vision-language models · COCO captioning Karpathy test | CIDEr: 137.6 -> 140.6 | Helped |
| Replace VIVO's vision model with VinVL | VIVO · NoCaps test overall | CIDEr: 86.6 -> 92.5 | Helped |
| Use the 3-way contrastive loss rather than the Oscar contrastive loss (ablation) | Oscar+ pre-training · VQA vqa-dev | accuracy: 69.8 -> 69.8 | No clear effect |
| Use the 3-way contrastive loss rather than the Oscar contrastive loss (ablation) | Oscar+ pre-training · COCO-IR 1K test | R@1: 73.9 -> 78.3 | Helped |
| Pre-train on four detection datasets before VG fine-tuning (ablation) | R50-C4 visual model · VQA vqa-dev | accuracy: 67.86 -> 68.39 | Helped |
| Increase the visual backbone from R50-C4 to X152-C4 with four-dataset pre-training and VG fine-tuning (ablation) | VinVL visual model · VQA vqa-dev | accuracy: 68.39 -> 71.34 | Helped |
| Add attribute training for 524 attribute classes (ablation) | R50-C4 trained on VG · VQA vqa-dev | accuracy: 66.51 -> 67.86 | Helped |
| Expand the object vocabulary from 317 classes to 1594 VG classes (ablation) | R50-C4 visual model · VQA vqa-dev | accuracy: 64.25 -> 66.51 | Helped |
| Replace Anderson R101-C4 weights with VinVL X152-C4 weights (ablation) | VQA model using Anderson region proposals · VQA vqa-dev | accuracy: 68.52 -> 70.25 | Helped |
| Use VinVL-proposed regions instead of Anderson-proposed regions (ablation) | VQA model using Anderson R101-C4 weights · VQA vqa-dev | accuracy: 68.52 -> 69.05 | Helped |
Closest papers by meaning
- Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · similarity 0.92)
- Synthesize Diagnose and Optimize: Towards Fine-Grained Vision-Language Understanding (CVPR 2024 · similarity 0.90)
- Seeing Clearly, Reasoning Confidently: Plug-and-Play Remedies for Vision Language Model Blindness (CVPR 2026 · similarity 0.90)
- Revisiting Few-Shot Object Detection with Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation (ICLR 2026 · similarity 0.89)
- VL-SAM-V2: Open-World Object Detection with General and Specific Query Fusion (NeurIPS 2025 · similarity 0.89)
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.