H2VU-Benchmark: A Comprehensive Benchmark for Hierarchical Holistic Video Understanding
👀 H²VU-Benchmark Overview
H²VU-Benchmark is designed to comprehensively assess the capabilities of video understanding models, particularly in real-world scenarios. It addresses limitations of existing benchmarks by focusing on extended video durations, advanced task complexity, and diversified real-world data.
Key Features
- Three-Tier Hierarchical Competency Classification: (L-1 to L-3) with 10,183 evaluation tasks covering a broad spectrum of diverse data.
- Two Main Categories:
- Offline General Video: Employs common perception and reasoning tasks, along with novel tasks focusing on countercommonsense comprehension and trajectory tracking. (27 evaluation task types)
- Online Streaming Video: Utilizes standard perception and reasoning tasks. (20 evaluation task types)
Key Differentiators from Existing Benchmarks
Our work distinguishes itself through three key features:
- Extended Video Duration:
- Encompasses a diverse range from a few seconds to 1.5 hours, significantly expanding the temporal scope.
- Evaluates models' ability to capture short-term dynamics and model long-term dependencies.
- Advanced Task Complexity:
- Builds on traditional perceptual and reasoning tasks with the introduction of two new modules:
- Counterfactual Reasoning: Assesses vision-oriented understanding through tasks that defy common sense (e.g., implausible causal relationships).
- Trajectory State Tracking: Evaluates the ability to track and predict the states and trajectories of targets in complex dynamic scenes.
- Builds on traditional perceptual and reasoning tasks with the introduction of two new modules:
- Diversified Real-World Data:
- Incorporates first-person streaming video data to better simulate real-world streaming data processing needs.
- Explores multimodal models' performance in understanding first-person streaming video, crucial for AI agents functioning as real-world assistants or autonomous agents.