README.md

metadata

license: apache-2.0
task_categories:
  - video-text-to-text
tags:
  - video-understanding
  - temporal-consistency
  - benchmark

TOC-Bench: A Temporal Object Consistency Benchmark for Video Large Language Models

Paper | GitHub Repository

TOC-Bench is a diagnostic benchmark designed to evaluate temporal object consistency in Video Large Language Models (Video-LLMs). While models often excel at general video understanding, TOC-Bench specifically tests their ability to maintain the identity, state, and continuity of objects across challenges such as occlusion, disappearance, reappearance, and state transitions.

Benchmark Details

• Size: 2,323 high-quality, human-verified QA pairs over 1,951 videos.
• Grounded reasoning: Each item is object-track grounded, linking subjects to per-frame trajectories and structured temporal event timelines.
• Dimensions: Covers 10 diagnostic dimensions, including event counting, event ordering, identity-sensitive reasoning, and hallucination-aware verification.
• Filtering: Employs a three-layer temporal-necessity filtering protocol to ensure that questions require temporal visual evidence rather than language priors or single-frame shortcuts.

Evaluation

The official repository provides a standardized evaluation harness to score model predictions.

Running Inference

To run a model on the benchmark using eval_runner.py:

python eval_runner.py \
  --bench <REPAIRED_BENCHMARK_JSON> \
  --videos-root <VIDEO_ROOT> \
  --video-registry <VIDEO_REGISTRY_JSON> \
  --model <MODEL_KEY> \
  --frames 16 \
  --concurrency 4 \
  --out predictions.jsonl

Computing Metrics

After generating predictions, you can compute accuracy and diagnostic metrics using:

python compute_metrics.py \
  --bench <REPAIRED_BENCHMARK_JSON> \
  --preds predictions.jsonl \
  --save-summary summary.json

Citation

If you find this benchmark useful in your research, please cite the following paper:

@article{tocbench2024,
  title={TOC-Bench: A Temporal Object Consistency Benchmark for Video Large Language Models},
  author={...},
  journal={arXiv preprint arXiv:2605.09904},
  year={2024}
}