Files changed (1) hide show
  1. README.md +82 -0
README.md ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ task_categories:
4
+ - video-text-to-text
5
+ - visual-question-answering
6
+ language: [en]
7
+ tags: [video, temporal-grounding, one-to-many, benchmark, evaluation, multimodal, mllm]
8
+ pretty_name: OMTG Bench
9
+ size_categories: [n<1K]
10
+ ---
11
+
12
+ # OMTG Bench: A Benchmark for One-to-Many Temporal Grounding
13
+
14
+ OMTG Bench is the first comprehensive benchmark tailored for One-to-Many Temporal Grounding,
15
+ introduced in **"Towards One-to-Many Temporal Grounding"** (ICML 2026, under review).
16
+
17
+ ## Dataset Summary
18
+
19
+ | Item | Value |
20
+ |---|---|
21
+ | # Samples | 340 |
22
+ | Annotation | Manually curated, expert-verified |
23
+ | Consistency | > 90% |
24
+ | Source video pools | Charades, ActivityNet, QVHighlights, VTimeLLM, Moment-10M (test only) |
25
+ | Overlap with training | None |
26
+ | Domains | Sports, cooking, news, etc. |
27
+
28
+ ### Distribution
29
+ | Property | Value |
30
+ |---|---|
31
+ | GT segments per query | 2 – 20 |
32
+ | 2–3 segments | 62.2% |
33
+ | > 6 segments | 15% |
34
+ | Video duration | 21 s – 17 min (avg 221.6 s) |
35
+
36
+ ## Evaluation Metrics
37
+ | Metric | Description |
38
+ |---|---|
39
+ | tIoU | Average temporal IoU over unions |
40
+ | C-Acc | Count Accuracy (cardinality correctness) |
41
+ | tF1@ξ | Temporal F1 at IoU ξ ∈ {0.3, 0.5, 0.7} via optimal bipartite matching |
42
+ | **EtF1** | Effective Temporal F1 — strictly penalizes wrong cardinality (primary metric) |
43
+
44
+ ## Baseline Results (OMTG Bench)
45
+ | Model | C-Acc | tF1@0.3 | tF1@0.5 | tF1@0.7 | tIoU | EtF1 |
46
+ |---|---|---|---|---|---|---|
47
+ | Qwen2.5-VL-3B | 0.00 | 15.17 | 7.01 | 2.86 | 11.60 | — |
48
+ | Gemini-3-Pro | 30.63 | 58.30 | 47.75 | 29.89 | 47.63 | 21.30 |
49
+ | Gemini-2.5-Pro | 50.94 | 55.72 | 43.57 | 27.97 | 43.24 | 27.80 |
50
+ | Seed-1.8 | 38.12 | 67.13 | 54.67 | 38.79 | 56.81 | 28.04 |
51
+ | **OMTG-4B (ours)** | — | — | — | — | — | **43.65** |
52
+
53
+ ## Dataset Structure
54
+ ```
55
+ omtg_bench/
56
+ ├── OMTGBench.tsv # 340 samples
57
+ └── videos.zip # ≈3.74 GB
58
+ ```
59
+
60
+ ### Data Fields
61
+ | Field | Type | Description |
62
+ |---|---|---|
63
+ | video_id | string | Video identifier |
64
+ | duration | float | Video length (s) |
65
+ | query | string | Natural-language query |
66
+ | segments | list[[float,float]] | GT `[start,end]` intervals |
67
+ | source | string | Charades / ActivityNet / QVHighlights / VTimeLLM / Moment-10M |
68
+ | domain | string | Topical domain |
69
+
70
+ ## Usage
71
+ ```python
72
+ from datasets import load_dataset
73
+ bench = load_dataset("insomnia7/omtg_bench", split="train")
74
+ ```
75
+
76
+ ## Companion Dataset
77
+ Training dataset: [insomnia7/omtg56k](https://huggingface.co/datasets/insomnia7/omtg56k) (46k SFT + 10k RL).
78
+
79
+ ## License
80
+ CC BY-NC 4.0 — non-commercial research only.
81
+
82
+ ```