File size: 7,696 Bytes
a29e644
 
 
 
 
565eb69
 
 
 
 
 
 
 
 
 
a29e644
565eb69
 
a29e644
 
 
 
565eb69
 
 
a29e644
 
 
 
 
565eb69
0f74cf1
 
 
 
 
565eb69
 
 
a29e644
0f74cf1
 
a29e644
 
0f74cf1
 
 
 
565eb69
 
 
 
 
 
a29e644
 
565eb69
 
 
a29e644
 
565eb69
 
 
 
 
 
 
 
 
 
a29e644
 
 
565eb69
a29e644
565eb69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a29e644
 
565eb69
 
 
a29e644
 
0f74cf1
 
a29e644
 
 
565eb69
a29e644
 
565eb69
a29e644
565eb69
a29e644
 
0f74cf1
565eb69
 
 
 
 
 
 
 
 
 
0f74cf1
565eb69
0f74cf1
 
 
 
565eb69
 
 
 
 
 
 
 
 
 
 
 
 
 
a29e644
565eb69
 
a29e644
 
 
565eb69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
---
license: cc-by-nc-4.0
task_categories:
  - video-text-to-text
  - visual-question-answering
language:
  - en
tags:
  - video
  - temporal-grounding
  - one-to-many
  - benchmark
  - evaluation
  - multimodal
  - mllm
pretty_name: OMTG Bench
size_categories:
  - n<1K
---

# OMTG Bench: A Benchmark for One-to-Many Temporal Grounding

**OMTG Bench** is the first comprehensive benchmark tailored for the **One-to-Many Temporal Grounding (OMTG)** task, introduced in the paper **"Towards One-to-Many Temporal Grounding"** (ICML 2026, under review).

Unlike conventional temporal grounding benchmarks that assume a one-to-one mapping between a query and a single temporal segment, OMTG Bench evaluates a model's ability to retrieve **all** disjoint temporal segments in a video that correspond to a single textual query — a setting that is ubiquitous in real-world video content but on which state-of-the-art MLLMs (including Gemini-2.5-Pro, Gemini-3-Pro, Seed-1.8, and Qwen2.5-VL / Qwen3-VL families) exhibit a critical capability gap.

## Dataset Summary

| Item | Value |
|---|---|
| Task | One-to-Many Temporal Grounding (OMTG) |
| Release files | `OMTGBench.tsv`, `videos.zip` |
| # Samples | 320 |
| # Unique videos | 287 |
| TSV columns | `id`, `video`, `question`, `answer` |
| Answer format | List of `[start, end]` intervals in seconds |
| Language | English |

### Key Distribution Statistics

The following statistics are computed directly from `OMTGBench.tsv`.

| Property | Value |
|---|---|
| Ground-truth segments per query | 2 - 20 |
| Average segments per query | 3.67 |
| 2–3 segments per query | 62.19% of samples |
| > 6 segments per query | 9.69% of samples |

This distribution makes OMTG Bench a severe test of both **event cardinality perception** and **fine-grained temporal localization**, covering short clips as well as long-form narratives.

## Supported Tasks

- **One-to-Many Temporal Grounding** — given a video `V` and a textual query `Q`, predict the set `{[s₁, e₁], [s₂, e₂], …, [sₖ, eₖ]}` of all disjoint intervals where `Q` occurs.

## Evaluation Metrics

The benchmark introduces a rigorous metric suite that jointly measures precision, recall, and event-count correctness:

| Metric | Description |
|---|---|
| **tIoU** | Average temporal IoU over unions of predicted and ground-truth intervals |
| **C-Acc** | Count Accuracy — the percentage of samples whose predicted cardinality equals the ground-truth cardinality |
| **tF1@ξ** | Temporal F1-Score at IoU threshold `ξ ∈ {0.3, 0.5, 0.7}` under optimal bipartite matching |
| **EtF1** | **Effective Temporal F1** — assigns zero credit to any sample with incorrect predicted cardinality, strictly penalizing under- and over-retrieval and hallucinations |

EtF1 is the recommended primary metric, as it couples instance-level precision–recall with event-count correctness.

## Baseline Results on OMTG Bench

A comprehensive assessment of representative open-source and proprietary MLLMs. All numbers are percentages (%).

| Model | C-Acc | tF1@0.3 | tF1@0.5 | tF1@0.7 | tIoU | EtF1 |
|---|---|---|---|---|---|---|
| **Proprietary Models** | | | | | | |
| Seed-1.8 | 38.12 | 67.13 | 54.67 | 38.79 | 56.81 | 28.04 |
| Gemini-2.5-Pro | 50.94 | 55.72 | 43.57 | 27.97 | 43.24 | 27.80 |
| Gemini-3-Pro | 30.63 | 58.30 | 47.75 | 29.89 | 47.63 | 21.30 |
| **Open-Source General MLLMs** | | | | | | |
| Qwen2.5-VL-3B | 0.00 | 15.17 | 7.01 | 2.86 | 11.60 | 0.00 |
| Qwen2.5-VL-7B | 0.00 | 21.04 | 12.08 | 7.14 | 20.35 | 0.00 |
| Qwen2.5-VL-32B | 0.00 | 16.81 | 9.66 | 4.76 | 18.32 | 0.00 |
| Qwen2.5-VL-72B | 0.00 | 21.16 | 12.20 | 6.88 | 20.02 | 0.00 |
| Qwen3-VL-4B | 0.31 | 37.07 | 26.75 | 17.93 | 30.42 | 0.21 |
| Qwen3-VL-8B | 0.00 | 37.73 | 27.02 | 18.70 | 30.62 | 0.00 |
| Qwen3-VL-30B | 0.00 | 37.03 | 25.98 | 17.52 | 32.36 | 0.00 |
| Qwen3-VL-235B | 0.31 | 34.66 | 25.25 | 16.45 | 25.56 | 0.21 |
| **Temporal-Grounding Experts** | | | | | | |
| VideoChat-R1-7B | 0.00 | 32.07 | 19.70 | 10.42 | 24.93 | 0.00 |
| VideoChat-R1.5-7B | 0.31 | 28.41 | 15.53 | 9.85 | 27.96 | 0.10 |
| Time-R1-7B | 0.00 | 28.94 | 18.73 | 10.00 | 24.11 | 0.00 |
| UniTime | 0.00 | 35.27 | 30.15 | 23.58 | 37.12 | 0.00 |
| Timelens-8B | 0.00 | 39.14 | 32.76 | 22.58 | 32.38 | 0.00 |
| **Ours** | | | | | | |
| **OMTG-4B** | **55.63** | **73.46** | **65.40** | **48.96** | **61.24** | **43.65** |

### Performance Gain of OMTG-4B across Training Stages

| Stage | C-Acc | tF1@0.3 | tF1@0.5 | tF1@0.7 | tIoU | EtF1 |
|---|---|---|---|---|---|---|
| Base (Qwen3-VL-4B) | 0.31 | 37.07 | 26.75 | 17.93 | 30.42 | 0.21 |
| + SFT (on OMTG-56K) | 44.06 | 69.57 | 61.23 | 45.63 | 56.94 | 34.81 |
| + RL (GRPO) | **55.63** | **73.46** | **65.40** | **48.96** | **61.24** | **43.65** |

Key observations:

- Standard open-source models (Qwen2.5-VL series) yield **0% C-Acc / 0% EtF1**, failing to capture the one-to-many complexity.
- Even the newer Qwen3-VL family (4B → 235B) barely exceeds 0% on cardinality-sensitive metrics.
- Advanced proprietary models (Seed-1.8, Gemini-2.5/3-Pro) reach only **21–28% EtF1**.
- Our **OMTG-4B** achieves a new state of the art at **43.65% EtF1**, outperforming the best proprietary baseline (Seed-1.8) by **+15.61 EtF1**.

## Dataset Structure

### Files

```
omtg_bench/
├── OMTGBench.tsv   # 320 annotated samples (72 kB)
└── videos.zip      # corresponding source videos
```

## Usage

```python
from datasets import load_dataset

bench = load_dataset("insomnia7/omtg_bench", split="train")
print(bench[0])
```

Videos can be extracted from `videos.zip` and matched via the `video` column.

### Recommended Evaluation Protocol

1. For each sample, prompt the model with the video and the textual query.
2. Parse the model output into a set of `[start, end]` intervals (seconds).
3. Compute **C-Acc**, **tF1@{0.3, 0.5, 0.7}**, **tIoU**, and **EtF1** via optimal bipartite matching against the ground-truth segments.
4. Report **EtF1** as the primary metric.

## Construction & Quality Control

This repository release contains a tab-separated benchmark file (`OMTGBench.tsv`) and an accompanying video archive (`videos.zip`). Each row in `OMTGBench.tsv` contains:

- `id`: sample identifier.
- `video`: video filename in `videos.zip`.
- `question`: textual query for one-to-many temporal grounding.
- `answer`: the full list of ground-truth `[start, end]` intervals in seconds.

## Intended Use

- Evaluating MLLMs on one-to-many video temporal grounding.
- Studying event cardinality perception and hallucination in video-language models.
- Diagnosing over- and under-retrieval behaviors in long-form video understanding.

## Limitations

- Source videos inherit biases and licenses from the original datasets.
- Domain coverage, while diverse, is not exhaustive; extremely long videos (> 20 min) are underrepresented.
- Manual annotations, although expert-verified, may contain rare residual noise.
- Intended for non-commercial research only.

## Companion Dataset

- Training dataset: [insomnia7/omtg56k](https://huggingface.co/datasets/insomnia7/omtg56k) — 56k high-quality instruction-tuning samples (46k SFT + 10k RL).

## License

Released under **CC BY-NC 4.0** for non-commercial research use. Source videos remain subject to their original dataset licenses.

## Citation

```bibtex
@inproceedings{omtg2026,
  title     = {Towards One-to-Many Temporal Grounding},
  author    = {Anonymous Authors},
  booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
  year      = {2026},
  note      = {Under review}
}
```

## Contact

For questions or issues, please open a discussion on the dataset's Hugging Face page.