FineBench: Benchmarking and Enhancing Vision-Language Models for Fine-grained Human Activity Understanding
Researchers have introduced FineBench, a new benchmark designed to evaluate the fine-grained human activity understanding capabilities of vision-language models (VLMs). The benchmark includes nearly 200,000 question-answer pairs across 64 long-form videos, focusing on detailed actions and interactions. Evaluations showed that while proprietary models like GPT-5 performed adequately, open-source VLMs struggled with spatial reasoning and subtle movement distinctions. To address these limitations, the team also proposed FineAgent, a framework that enhances VLMs using a localizer and descriptor, demonstrating improved performance on FineBench. AI
IMPACT Establishes a new standard for evaluating VLM's nuanced human activity understanding, potentially driving development of more capable models.