VCBench: 4,574 Video Clips for Object and Event Counting Tasks

VCBench contains 4,574 clipped video segments totaling approximately 80 GB, developed by buaaplay and last updated in March 2026. The collection is organized into 8 subcategories specifically designed to evaluate spatial-temporal state maintenance in video understanding models. It focuses on two primary task types: object counting and event counting.

Use Cases

• Evaluating spatial-temporal reasoning by counting specific objects within video segments
• Benchmarking video classification models on event-based counting subcategories
• Training visual question answering systems to handle numerical queries about temporal sequences

Strengths

• 4,574 high-quality MP4 video clips
• 8 specialized subcategories for granular counting evaluation
• MIT licensed for open research and commercial use

Limitations

• Large storage footprint of ~80 GB may pose accessibility challenges for users with limited bandwidth
• Narrow focus on counting tasks may not generalize to broader video understanding benchmarks

Provenance

Source

buaaplay (Video Counting Benchmark)

Freshness

Last updated March 12, 2026.