Research map / Vision-language and multimodal / Visual question answering
Contrast and Classify: Training Robust VQA Models
Yash Kant, Abhinav Moudgil, Dhruv Batra, Devi Parikh, Harsh Agrawal
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 ConClaT alternate contrastive+CE training with BT data | MMT baseline · VQA-Rephrasings | Consensus Score CS(4) (%): +1.63% | Improved |
| add ConClaT alternate contrastive+CE training with BT data | MMT baseline · VQA 2.0 val | VQA accuracy (%): 66.31 -> 66.98 | Improved |
| add ConClaT with VQG-generated paraphrases | MMT + VQG data · VQA-Rephrasings | Consensus Score CS(4) (%): +0.46% | Improved |
| add back-translated paraphrase data with CE only | MMT · VQA-Rephrasings | Consensus Score CS(4) (%): +0.54% | Improved |
| train jointly instead of alternately (ablation) | ConClaT · VQA-Rephrasings | Consensus Score CS(4) (%): -0.36% | Hurt |
| use pretrain-finetune instead of alternate training (ablation) | ConClaT · VQA-Rephrasings | Consensus Score CS(4) (%): -4.79% | Hurt |
| replace L_SSCL with dynamic-margin triplet loss (ablation) | ConClaT · VQA-Rephrasings | Consensus Score CS(4) (%): 53.99 -> 53.10 | Hurt |
| add scaling factor to supervised contrastive loss (ablation) | ConClaT · VQA-Rephrasings | Consensus Score CS(4) (%): 53.36 -> 53.99 | Improved |
| use dynamic scaling instead of constant scaling (ablation) | ConClaT · VQA-Rephrasings | Consensus Score CS(4) (%): 53.99 vs 53.92 | No clear effect |
| use curated question+image+random negatives instead of random negatives (ablation) | ConClaT · VQA-Rephrasings | Consensus Score CS(4) (%): 53.42 -> 53.99 | Improved |
| use both image-type and question-type negatives rather than one type (ablation) | ConClaT · VQA-Rephrasings | Consensus Score CS(4) (%): 53.77 -> 53.99 | Improved |
| raise joint-loss weight beta from 0.5 to 0.9 (ablation) | Joint L_SSCL+L_CE · VQA-Rephrasings | Consensus Score CS(4) (%): 53.63 -> 40.68 | Hurt |
Closest papers by meaning
- Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering (CVPR 2018 · similarity 0.93)
- Plenty is Plague: Fine-Grained Learning for Visual Question Answering (TPAMI 2019 · similarity 0.93)
- Cycle-Consistency for Robust Visual Question Answering (CVPR 2019 · similarity 0.93)
- Robust Visual Question Answering: Datasets, Methods, and Future Challenges (TPAMI 2024 · similarity 0.92)
- Self-Critical Reasoning for Robust Visual Question Answering (NeurIPS 2019 · similarity 0.92)
- FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.91)
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.
