File size: 4,854 Bytes
3e0a629
 
 
f5e7212
3e0a629
 
 
3d8e168
3e0a629
 
 
 
 
 
 
31a402f
 
3e0a629
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31a402f
 
 
 
 
2d18b98
 
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
2d18b98
31a402f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3e0a629
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31a402f
3e0a629
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
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
---
license: apache-2.0
task_categories:
  - object-detection
language:
  - en
tags:
  - referring-expression-comprehension
  - visual-grounding
  - refcocog
  - fine-grained-evaluation
size_categories:
  - 1K<n<10K
---

# RefCOCOg-UMD Test Partitioning

This dataset provides **fine-grained evaluation splits** for the RefCOCOg-UMD test set, designed for Referring Expression Comprehension (REC) tasks. Each sample is partitioned by **difficulty level** and **referential type** to enable more detailed performance analysis.

## Dataset Description

- **Total Samples:** 9,482 (by difficulty) / 19,334 (by referential type, with overlaps)
- **Source:** RefCOCOg-UMD test set
- **Format:** PyTorch `.pth` files

## Data Structure

Each `.pth` file contains a list of tuples with the following structure:

```python
import torch

data = torch.load("Easy.pth")
for item in data:
    img_name, hw_dict, bbox, phrase, obj_mask = item
    # img_name: str - Image filename
    # hw_dict: dict - Height and width info {"height": int, "width": int}
    # bbox: list - Bounding box [x1, y1, x2, y2]
    # phrase: str - Referring expression text
    # obj_mask: tensor/array - Object segmentation mask
```

## File Organization

### By Difficulty Level
- `Easy.pth` - Simple referring expressions (4,109 samples)
- `Medium.pth` - Moderate complexity (4,517 samples)
- `Hard.pth` - Complex expressions (856 samples)

### By Referential Type × Difficulty
- `{Type}_Easy.pth`, `{Type}_Medium.pth`, `{Type}_Hard.pth`
- Types: `Attribute`, `Relation`, `Logic`, `Ambiguity`, `Perspective`

## Statistics for Each File

<table border="1" style="border-collapse: collapse;">
  <tr>
    <th>Filename</th>
    <th>Avg. Length</th>
    <th>Samples</th>
  </tr>
  <tr>
    <td>Attribute_Easy.pth</td>
    <td>35.92</td>
    <td>3,076</td>
  </tr>
  <tr>
    <td>Attribute_Medium.pth</td>
    <td>48.74</td>
    <td>3,431</td>
  </tr>
  <tr>
    <td>Attribute_Hard.pth</td>
    <td>50.90</td>
    <td>563</td>
  </tr>
  <tr>
    <td>Relation_Easy.pth</td>
    <td>37.86</td>
    <td>2,673</td>
  </tr>
  <tr>
    <td>Relation_Medium.pth</td>
    <td>47.51</td>
    <td>4,092</td>
  </tr>
  <tr>
    <td>Relation_Hard.pth</td>
    <td>51.95</td>
    <td>709</td>
  </tr>
  <tr>
    <td>Logic_Easy.pth</td>
    <td>49.92</td>
    <td>172</td>
  </tr>
  <tr>
    <td>Logic_Medium.pth</td>
    <td>53.27</td>
    <td>376</td>
  </tr>
  <tr>
    <td>Logic_Hard.pth</td>
    <td>61.68</td>
    <td>118</td>
  </tr>
  <tr>
    <td>Ambiguity_Easy.pth</td>
    <td>32.35</td>
    <td>174</td>
  </tr>
  <tr>
    <td>Ambiguity_Medium.pth</td>
    <td>42.62</td>
    <td>1,659</td>
  </tr>
  <tr>
    <td>Ambiguity_Hard.pth</td>
    <td>45.64</td>
    <td>783</td>
  </tr>
  <tr>
    <td>Perspective_Easy.pth</td>
    <td>30.07</td>
    <td>299</td>
  </tr>
  <tr>
    <td>Perspective_Medium.pth</td>
    <td>44.29</td>
    <td>955</td>
  </tr>
  <tr>
    <td>Perspective_Hard.pth</td>
    <td>52.14</td>
    <td>254</td>
  </tr>
  <tr>
    <td>Easy.pth</td>
    <td>33.85</td>
    <td>4,109</td>
  </tr>
  <tr>
    <td>Medium.pth</td>
    <td>46.19</td>
    <td>4,517</td>
  </tr>
  <tr>
    <td>Hard.pth</td>
    <td>47.87</td>
    <td>856</td>
  </tr>
</table>


## Statistics for Difficulty Levels and Referential Types
**STATISTICS OF THE FINE-GRAINED EVALUATION SPLITS ON REFCOCOG-UMD TEST SET. EACH SAMPLE IS ASSIGNED ONE DIFFICULTY LEVEL, WHILE REFERENTIAL CATEGORIES MAY OVERLAP ACROSS SAMPLES.**

<table border="1" style="border-collapse: collapse;">
  <tr>
    <th>Category</th>
    <th>Subset</th>
    <th>Avg. Length</th>
    <th>Samples</th>
    <th>Total</th>
  </tr>
  <tr>
    <td rowspan="3">Difficulty</td>
    <td>Easy</td>
    <td>33.85</td>
    <td>4,109</td>
    <td rowspan="3">9,482</td>
  </tr>
  <tr>
    <td>Medium</td>
    <td>46.19</td>
    <td>4,517</td>
  </tr>
  <tr>
    <td>Hard</td>
    <td>47.87</td>
    <td>856</td>
  </tr>
  <tr>
    <td rowspan="5">Referential Type</td>
    <td>Attribute</td>
    <td>43.33</td>
    <td>7,070</td>
    <td rowspan="5">19,334</td>
  </tr>
  <tr>
    <td>Relation</td>
    <td>44.48</td>
    <td>7,474</td>
  </tr>
  <tr>
    <td>Logic</td>
    <td>53.89</td>
    <td>666</td>
  </tr>
  <tr>
    <td>Ambiguity</td>
    <td>42.84</td>
    <td>2,616</td>
  </tr>
  <tr>
    <td>Perspective</td>
    <td>42.79</td>
    <td>1,508</td>
  </tr>
</table>

## Usage

```python
from huggingface_hub import hf_hub_download
import torch

# Download a specific file
file_path = hf_hub_download(
    repo_id="marloweee/BARE_grefumd_test_partitioning",
    filename="Easy.pth",
    repo_type="dataset"
)

# Load and iterate
data = torch.load(file_path)
for img_name, hw_dict, bbox, phrase, obj_mask in data:
    print(f"Image: {img_name}, Query: {phrase}")
```

## License

This dataset is released under the Apache 2.0 License.