Zhong
commited on
Update README.md
Browse files
README.md
CHANGED
@@ -64,11 +64,11 @@ Referring Expression Generation (REG)—the task of producing a concise and unam
|
|
64 |
1. **Data leakage in RefCOCO/RefCOCO+**, which raises concerns about evaluation contamination, especially for VLMs trained on MSCOCO.
|
65 |
2. **Lack of spoken data**, despite the fact that real-world referring is often **real-time** and **spontaneous**, unlike written language, which benefits from planning and revision.
|
66 |
|
67 |
-
|
68 |
|
69 |
**Key features:**
|
70 |
- 1,485 real-world object instances, equally distributed across **COCO** (744) and **non-COCO** (741) classes.
|
71 |
-
- Includes **single presence** and **co-
|
72 |
- Each instance annotated with **3 written** and **2 spoken** human referring expressions.
|
73 |
|
74 |
|
@@ -78,12 +78,12 @@ Using RefOI, we evaluate several state-of-the-art VLMs and uncover **three tiers
|
|
78 |
- **Redundancy**: Models include excessive or irrelevant details, violating principles of informativeness and efficiency.
|
79 |
- **Misalignment**: Model preferences diverge from human pragmatics, favoring visual complexity over minimal spatial cues.
|
80 |
|
81 |
-
Our results also highlight the inadequacy of standard automatic metrics (e.g., BLEU, CIDEr) and listener-based scores (e.g., REC), which fail to capture these pragmatic shortcomings—emphasizing the need for more cognitively grounded evaluation protocols.
|
82 |
-
|
83 |

|
84 |
|
85 |
-
## Dataset Structure
|
86 |
|
|
|
|
|
|
|
87 |
Each entry in the dataset contains the following fields:
|
88 |
|
89 |
- `image`: The original image file.
|
@@ -96,9 +96,12 @@ Each entry in the dataset contains the following fields:
|
|
96 |
- `written_descriptions`: Three human‑typed referring expressions.
|
97 |
- `spoken_descriptions`: Two human‑spoken expressions (transcribed and optionally corrected by annotators).
|
98 |
|
99 |
-
|
100 |
-
`single_presence
|
101 |
-
|
|
|
|
|
|
|
102 |
|
103 |
|
104 |
## Usage
|
@@ -106,13 +109,13 @@ Each entry in the dataset contains the following fields:
|
|
106 |
```python
|
107 |
from datasets import load_dataset
|
108 |
|
109 |
-
#
|
110 |
-
|
111 |
-
#
|
112 |
-
|
113 |
|
114 |
-
print(
|
115 |
-
print(
|
116 |
```
|
117 |
|
118 |
|
@@ -149,7 +152,7 @@ Model performance under different **Instr.** (Instruction) settings: **Dft.** (D
|
|
149 |
|
150 |
## Recommended Use of Our Dataset
|
151 |
|
152 |
-
The `RefOI` dataset is designed for fine-grained REG/REC analysis. It distinguishes between **COCO** and **non-COCO classes**, and between scenes with **single presence vs. co-
|
153 |
We encourage users to leverage these distinctions for deeper insights and invite community contributions to expand non-COCO annotations.
|
154 |
|
155 |
|
|
|
64 |
1. **Data leakage in RefCOCO/RefCOCO+**, which raises concerns about evaluation contamination, especially for VLMs trained on MSCOCO.
|
65 |
2. **Lack of spoken data**, despite the fact that real-world referring is often **real-time** and **spontaneous**, unlike written language, which benefits from planning and revision.
|
66 |
|
67 |
+
To address these gaps, we introduce **RefOI**, a curated dataset built from the [OpenImages V7](https://storage.googleapis.com/openimages/web/index.html) Instance Segmentation validation set.
|
68 |
|
69 |
**Key features:**
|
70 |
- 1,485 real-world object instances, equally distributed across **COCO** (744) and **non-COCO** (741) classes.
|
71 |
+
- Includes **single presence** and **co-occurrence** images for each class.
|
72 |
- Each instance annotated with **3 written** and **2 spoken** human referring expressions.
|
73 |
|
74 |
|
|
|
78 |
- **Redundancy**: Models include excessive or irrelevant details, violating principles of informativeness and efficiency.
|
79 |
- **Misalignment**: Model preferences diverge from human pragmatics, favoring visual complexity over minimal spatial cues.
|
80 |
|
|
|
|
|
81 |

|
82 |
|
|
|
83 |
|
84 |
+
## ## Dataset Schema and Split
|
85 |
+
|
86 |
+
### Data Fields
|
87 |
Each entry in the dataset contains the following fields:
|
88 |
|
89 |
- `image`: The original image file.
|
|
|
96 |
- `written_descriptions`: Three human‑typed referring expressions.
|
97 |
- `spoken_descriptions`: Two human‑spoken expressions (transcribed and optionally corrected by annotators).
|
98 |
|
99 |
+
### Dataset Split
|
100 |
+
- `single_presence` (`co_occurrence = 1`):
|
101 |
+
Only one object of the target class appears (no same‑class distractors in the image).
|
102 |
+
|
103 |
+
- `co_occurrence` (`co_occurrence > 1`):
|
104 |
+
Multiple objects of the same class appear in the image, introducing potential referential ambiguity.
|
105 |
|
106 |
|
107 |
## Usage
|
|
|
109 |
```python
|
110 |
from datasets import load_dataset
|
111 |
|
112 |
+
# only one object of the class
|
113 |
+
ds_single = load_dataset("Seed42Lab/RefOI", split="single_presence")
|
114 |
+
# multiple objects of the class
|
115 |
+
ds_multi = load_dataset("Seed42Lab/RefOI", split="co_occurrence")
|
116 |
|
117 |
+
print(ds_single[0])
|
118 |
+
print(ds_multi[0])
|
119 |
```
|
120 |
|
121 |
|
|
|
152 |
|
153 |
## Recommended Use of Our Dataset
|
154 |
|
155 |
+
The `RefOI` dataset is designed for fine-grained REG/REC analysis. It distinguishes between **COCO** and **non-COCO classes**, and between scenes with **single presence vs. co-occurrence** of the same class.
|
156 |
We encourage users to leverage these distinctions for deeper insights and invite community contributions to expand non-COCO annotations.
|
157 |
|
158 |
|