Hugging Face's logo

Datasets:
Salesforce
/
ST-Evidence-Bench

Dataset card Files
xet
 Community
1
ST-Evidence-Bench
File size: 4,355 Bytes
1841309
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# ST-Evidence Benchmark Dataset

ST-Evidence is a comprehensive benchmark for evaluating Spatial-Temporal Evidence generation in video understanding. It contains two tasks: **Generation (Gen)** and **Multiple Choice Question (MCQ)**.

This was released for research purposes only, in support of the academic paper Evidence-Backed Video Question Answering.

## Dataset Overview

- **Total Videos**: ~1,300 videos at 6fps
- **Annotations**: Question-Answer pairs with temporal segments and spatial masks
- **Tasks**: Generation and MCQ
- **Domains**: Diverse video content

## Files Structure

```
ST-Evidence/
├── st_evidence_gen/              # Generation Task
│   ├── st_evidence_gen.csv       # 924KB - Annotations (entry_id, question, answer, segments, etc.)
│   ├── videos_6fps.tar.gz        # 8.3GB - Video files at 6fps
│   └── masks.tar.gz              # 560MB - Ground truth spatial masks
│
└── st_evidence_mcq/              # Multiple Choice Question Task
    ├── st_evidence_mcq.csv       # 313KB - MCQ annotations
    ├── mask_options.json         # 575KB - Mask options metadata
    ├── temp_options.json         # 679KB - Temporal options metadata
    └── options.tar.gz            # 1.5GB - Pre-rendered option masks (1,298 entries)
```

## Generation Task (st_evidence_gen)

### Data Format

**st_evidence_gen.csv** contains the following columns:
- `entry_id`: Unique identifier for each question
- `video_id`: Video identifier
- `video_path`: Relative path to video file
- `question`: Question text
- `candidates`: List of answer options (for reference)
- `answer`: Ground truth answer
- `segment`: Temporal evidence segments [[start1, end1], [start2, end2], ...]

### Usage

```python
import pandas as pd
import tarfile

# Load annotations
df = pd.read_csv('st_evidence_gen.csv')

# Extract videos
with tarfile.open('videos_6fps.tar.gz', 'r:gz') as tar:
    tar.extractall('videos_6fps/')

# Extract ground truth masks
with tarfile.open('masks.tar.gz', 'r:gz') as tar:
    tar.extractall('masks/')
```

### Evaluation Metrics

- **QA Accuracy**: Percentage of correct answers
- **Temporal IoU**: Intersection over Union for temporal segments
  - mIoU, TIoU@0.3, TIoU@0.5
- **Temporal IoP**: Intersection over Prediction
  - mIoP, TIoP@0.3, TIoP@0.5
- **Spatial Quality** (if masks generated):
  - J score (Jaccard/IoU)
  - F score (contour-based)
  - J&F score (average)

## MCQ Task (st_evidence_mcq)

### Data Format

**st_evidence_mcq.csv** contains:
- `entry_id`: Unique identifier
- `video_id`: Video identifier
- `video_path`: Path to video
- `question`: Question text
- `candidates`: Answer options
- `answer`: Correct answer
- `segment`: Temporal evidence
- `mask_options`: Reference to mask options
- `temp_options`: Reference to temporal options

**mask_options.json**: Contains spatial mask options for each question
**temp_options.json**: Contains temporal segment options for each question
**options.tar.gz**: Pre-rendered mask visualizations for options (1,298 entries)

### Usage

```python
import json
import pandas as pd

# Load MCQ annotations
df = pd.read_csv('st_evidence_mcq.csv')

# Load options
with open('mask_options.json', 'r') as f:
    mask_options = json.load(f)

with open('temp_options.json', 'r') as f:
    temp_options = json.load(f)

# Extract option masks
with tarfile.open('options.tar.gz', 'r:gz') as tar:
    tar.extractall('options/')
```

### Evaluation Metrics

Same as Generation task, but with multiple-choice format.

## Download & Setup

### Using HuggingFace Hub

```python
from huggingface_hub import snapshot_download

# Download entire dataset
snapshot_download(
    repo_id="Salesforce/ST-Evidence",
    repo_type="dataset",
    local_dir="./st_evidence_data"
)
```

### Manual Download

1. Download all files from this repository
2. Extract compressed files:
   ```bash
   tar -xzf videos_6fps.tar.gz
   tar -xzf masks.tar.gz
   tar -xzf options.tar.gz
   ```

## Citation

If you use this dataset, please cite:

```bibtex
@article{st-evidence2025,
  title={ST-Evidence: A Benchmark for Spatial-Temporal Evidence in Video Understanding},
  author={Wang, Shijie and others},
  year={2025}
}
```

## License

CC-BY-NC 4.0


## Version

- **Version**: 1.0
- **Release Date**: 2025-03-14
- **Total Size**: ~10.4 GB (compressed)