CoVeR-VQA improves multi-view driving scene question answering accuracy
GroundSight at GroundLM 2026 Shared Tasks: GoldenViewVQA
Computer Vision and Pattern Recognition
Summary
Answering questions about driving scenes requires not only finding the correct answer but also pinpointing which camera view shows the supporting details. The authors present CoVeR-VQA, a multi-step method that improves accuracy by verifying and correcting answers from AI models. Their approach uses several AI tools to double-check answers and where the evidence is found, raising accuracy significantly over initial AI guesses. They found that locating the right camera view remains the main challenge, showing the importance of careful evidence checking.
What this means in practice
- •For autonomous vehicle developers: Improve scene understanding by verifying which camera view supports visual answers to driving scenario questions.
- •For multimodal ai system builders: Use a multi-stage AI verification framework to boost precision in joint answer and evidence localization tasks.
Authors
Kun Wang, Yupeng Hu, Ruping Cao, Hao Liu, Zhiran Li, Qianlong Xiang, Harry Cheng
Abstract
GoldenViewVQA requires models to jointly answer driving-scene questions and identify the camera view containing the supporting visual evidence, making precise evidence localization as important as answer correctness. We present \textbf{CoVeR-VQA}, a training-free multi-stage verification and correction framework for grounded multi-view VQA. Starting from GPT-5.6 zero-shot predictions, CoVeR-VQA progressively applies view-specific verification with Gemini-3.6-Flash, prior-guided joint verification with Claude-Opus-5, and cross-split group-level verification that exploits semantically filtered question groups from shared multi-view scenes and validation-derived prior knowledge. On the official GoldenViewVQA test set, the four-stage CoVeR-VQA pipeline achieves 84.75\% Joint Accuracy, improving the GPT-5.6 zero-shot baseline by 13.56 percentage points, while reaching 94.92\% Answer Accuracy and 86.44\% View Accuracy. The final submitted run achieves 88.14\% Joint Accuracy after two additional evaluator-informed post-hoc corrections. Our analysis shows that supporting-view localization remains the primary source of residual errors, highlighting the importance of explicit evidence verification for reliable multi-view multimodal reasoning.