IC-738BLIP-2 ViT-G FlanT5XL achieves 0.70–0.87 zero-shot accuracy on binary concept presence queries, competitive on most datasets but weaker on fine-grained CUB-200

Aditya Chattopadhyay, Kwan Ho Ryan Chan, Rene Vidal

SourceBootstrapping Variational Information Pursuit with Large Language and Vision Models for Interpretable Image Classification

The paper evaluates BLIP-2 ViT-G FlanT5XL in a zero-shot setting, prompting it with 'question: is [concept] present in the image (yes/no)? answer:' to answer binary concept queries. It reaches 0.86 accuracy on ImageNet, 0.87 on Places365, 0.70 on CUB-200, 0.83 on CIFAR-10, and 0.87 on CIFAR-100. The authors note it is competitive with their 150M-parameter Concept-QA on CIFAR and ImageNet, considerably worse on CUB-200 (a fine-grained bird species dataset), and better on Places365. The alternative BLIP-2 variant (ViT-G OPT-2.7B) performs worse overall, at 0.53–0.86. The authors attribute the CUB-200 gap to the domain-specific nature of the dataset being too specialized for a general-purpose VQA model.

Evidence
correlational
Key metric
BLIP-2 ViT-G FlanT5XL accuracy: 0.86 (ImageNet), 0.87 (Places365), 0.70 (CUB-200), 0.83 (CIFAR-10), 0.87 (CIFAR-100); BLIP-2 ViT-G OPT-2.7B: 0.55, 0.76, 0.53, 0.73, 0.86
Caveat
Ground truth is self-annotated by the authors (2,500 image-concept pairs per dataset). The prompt format is specific to this paper's setup. The authors note BLIP-2 is computationally prohibitive for V-IP training on large datasets (~4,128 GPU hours per epoch on ImageNet).
Model
BLIP-2
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval], Places365 [eval], CUB-200-2011 / CUB200 [eval], CIFAR-10 [eval], CIFAR-100 [eval]
Related findings
IC-737, IC-739
Extraction
automatic-extraction