Open-ended VQA benchmarking of Vision-Language models by exploiting Classification datasets and their semantic hierarchy

11 Feb 2024  ·  Simon Ging, María A. Bravo, Thomas Brox ·

The evaluation of text-generative vision-language models is a challenging yet crucial endeavor. By addressing the limitations of existing Visual Question Answering (VQA) benchmarks and proposing innovative evaluation methodologies, our research seeks to advance our understanding of these models' capabilities. We propose a novel VQA benchmark based on well-known visual classification datasets which allows a granular evaluation of text-generative vision-language models and their comparison with discriminative vision-language models. To improve the assessment of coarse answers on fine-grained classification tasks, we suggest using the semantic hierarchy of the label space to ask automatically generated follow-up questions about the ground-truth category. Finally, we compare traditional NLP and LLM-based metrics for the problem of evaluating model predictions given ground-truth answers. We perform a human evaluation study upon which we base our decision on the final metric. We apply our benchmark to a suite of vision-language models and show a detailed comparison of their abilities on object, action, and attribute classification. Our contributions aim to lay the foundation for more precise and meaningful assessments, facilitating targeted progress in the exciting field of vision-language modeling.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Visual Question Answering (VQA) ActivityNet BLIP-2 T5 ClipMatch@1 53.39 # 1
ClipMatch@5 74.71 # 1
Contains 15.70 # 1
ExactMatch 7.07 # 1
Follow-up ClipMatch@1 62.02 # 1
Follow-up ClipMatch@5 75.13 # 1
Follow-up Contains 18.09 # 1
Follow-up ExactMatch 8.84 # 1
Visual Question Answering (VQA) COCO InstructBLIP Vicuna ClipMatch@1 59.58 # 1
ClipMatch@5 73.32 # 1
Contains 27.52 # 1
ExactMatch 26.50 # 1
Visual Question Answering (VQA) ImageNet BLIP-2 OPT Contains 35.49 # 1
ClipMatch@1 57.10 # 1
ClipMatch@5 77.24 # 1
Follow-up ClipMatch@1 67.22 # 1
Follow-up ClipMatch@5 83.54 # 1
Follow-up Contains 40.31 # 1
Follow-up ExactMatch 2.54 # 1
ExactMatch 0.87 # 1
Visual Question Answering (VQA) OVAD benchmark BLIP Contains w. Synonyms 45.70 # 1
ExactMatch w. Synonyms 36.99 # 1

Methods


No methods listed for this paper. Add relevant methods here