Support OMTG Benchmark

[insomniaaac]

Description

This PR adds support for the One-to-Many Temporal Grounding (OMTG) benchmark, as proposed in the paper Towards One-to-Many Temporal Grounding.

Unlike traditional temporal grounding tasks that assume a one-to-one mapping, OMTG requires the model to localize all disjoint video segments corresponding to a query.

Key Changes

• New Benchmark Support: Added OMTGBench dataset class.
• New Metrics: Implemented rigorous metrics for multi-instance retrieval:
  • C-Acc (Count Accuracy): Evaluates event cardinality perception.
  • EtF1 (Effective Temporal F1): The primary metric that penalizes incomplete retrieval.
  • tF1 (Temporal F1-Score).
• Evaluation Pipeline: Integrated the OMTG evaluation logic into the existing framework.

How to Use

Users can evaluate models on the OMTG benchmark using the following command:

python run.py --data OMTGBench --model Qwen3-VL-4B-Instruct --verbose

[FangXinyu-0913]

1. Consider add quick config in vlmeval/dataset/video_dataset_config.py for better usage
2. Please report OMTG Bench performance for representative models (using VLMEvalKit, the official repo, and paper results). Include environment details (transformers, torch, vllm/sglang, flash-attention, python) and specific configs (like nframe) used for these runs.
3. Please help fix the lint: https://github.com/open-compass/VLMEvalKit/actions/runs/21662694744/job/62492361066?pr=1427