Datasets:
Formats:
parquet
Languages:
English
Size:
1K - 10K
Tags:
referring-expression-comprehension
visual-grounding
text-object-grounding
ocr
scene-text
autonomous-driving
License:
Sync trimmed README
Browse files
README.md
CHANGED
|
@@ -77,95 +77,3 @@ CAM_FRONT, CAM_BACK_LEFT, CAM_BACK_RIGHT).
|
|
| 77 |
| `num_texts_used` | int | how many text instances were combined into this sample |
|
| 78 |
| `sample_type` | string | `single_text_single_object`, `multi_text_single_object`, or `disambiguated_same_category` |
|
| 79 |
| `attribute` | string (optional) | present only for `disambiguated_same_category` samples — a distinguishing visual/spatial attribute added because another object of the *same category* in the same image shares the same text (e.g. `"left blue"` in `"'23' written on left blue gate"`) |
|
| 80 |
-
|
| 81 |
-
### Sample types (paper's scenario taxonomy)
|
| 82 |
-
|
| 83 |
-
The paper defines four scene-text-to-object scenarios (Sec. 3.1); this
|
| 84 |
-
release covers three of them:
|
| 85 |
-
|
| 86 |
-
- **`single_text_single_object`** — paper's *Single-text → Single-object*:
|
| 87 |
-
one object, one piece of text.
|
| 88 |
-
- **`multi_text_single_object`** — paper's *Multi-text → Single-object*:
|
| 89 |
-
one object with multiple pieces of text on it; one sample is generated
|
| 90 |
-
per count *k* = 1..5 using the *k* spatially closest texts, so the same
|
| 91 |
-
object contributes several samples of increasing expression detail.
|
| 92 |
-
- **`disambiguated_same_category`** — paper's *Single-text → Multi-object*:
|
| 93 |
-
used when two or more objects of the **same category** in the same image
|
| 94 |
-
share the exact same text string (so category + text alone would be
|
| 95 |
-
ambiguous). Following the paper's attribute generation and
|
| 96 |
-
disambiguation process, a human-annotated distinguishing attribute
|
| 97 |
-
(color, material, shape, and/or left/right/top/center position) is
|
| 98 |
-
folded into the expression: `'{text}' written on {attribute} {category}`.
|
| 99 |
-
Cross-category sharing (e.g. a gate and a building both labeled `"303"`)
|
| 100 |
-
is **not** treated as ambiguous here, since the category name itself
|
| 101 |
-
already disambiguates those.
|
| 102 |
-
- *Multi-instance ↔ Multi-text* (the paper's fourth scenario, where
|
| 103 |
-
several objects each carry several text instances) is not yet
|
| 104 |
-
represented as its own sample type in this nuScenes release.
|
| 105 |
-
|
| 106 |
-
## Construction pipeline
|
| 107 |
-
|
| 108 |
-
This mirrors the paper's general data curation pipeline (Fig. 4: Data
|
| 109 |
-
Collection → Annotation Curation → Processing → Attribute Curation →
|
| 110 |
-
Quality Control → Data Taxonomy), instantiated for nuScenes using 3D box
|
| 111 |
-
projection in place of the open-vocabulary detectors/segmentation models
|
| 112 |
-
used for the OpenImages/COCO portions:
|
| 113 |
-
|
| 114 |
-
1. **OCR**: EasyOCR and PaddleOCR run independently on every camera frame.
|
| 115 |
-
A text box is kept only if both models agree exactly on the recognized
|
| 116 |
-
string, both confidences exceed 0.7, and the IoU between their two boxes
|
| 117 |
-
exceeds 0.7.
|
| 118 |
-
2. **Scene-aware temporal dedup**: nuScenes scenes are ~20s clips at 2Hz;
|
| 119 |
-
the same physical text is visible across many consecutive frames. Runs
|
| 120 |
-
of the same `(scene, camera, text)` are collapsed to one representative
|
| 121 |
-
frame per spatially-contiguous run.
|
| 122 |
-
3. **3D object association**: each nuScenes 3D object annotation is
|
| 123 |
-
projected into the camera image (ego/global frame transforms + camera
|
| 124 |
-
intrinsics); a text box is linked to an object if the object's projected
|
| 125 |
-
silhouette contains it — the dataset's analog of the paper's
|
| 126 |
-
detector/segmentation-based object candidates.
|
| 127 |
-
4. **Manual review**: every image with 2+ text detections was manually
|
| 128 |
-
reviewed and corrected (text/object boxes fixed, text-to-object links
|
| 129 |
-
verified) via a custom annotation tool.
|
| 130 |
-
5. **Categorization**: free-text object names were mapped to a fixed
|
| 131 |
-
category taxonomy (`vehicle.*`, `signage.*`, `structure.*`, `barrier.*`,
|
| 132 |
-
`misc.*`).
|
| 133 |
-
6. **REC sample construction**: samples generated per the scenarios above,
|
| 134 |
-
with every `combined_text_box` guaranteed to be fully contained in its
|
| 135 |
-
`obj_box`.
|
| 136 |
-
7. **Attribute generation and disambiguation**: same-category ties (a
|
| 137 |
-
shared text string across 2+ objects of the same category) were
|
| 138 |
-
manually annotated with a distinguishing attribute so a unique
|
| 139 |
-
referring expression is always possible, per the paper's Sec. 3.1
|
| 140 |
-
attribute-curation process.
|
| 141 |
-
|
| 142 |
-
## Source data & license
|
| 143 |
-
|
| 144 |
-
Images are camera frames from the [nuScenes](https://www.nuscenes.org/)
|
| 145 |
-
dataset (© nuTonomy / Motional), used under the nuScenes non-commercial
|
| 146 |
-
license. This derived annotation set (`annotations.json`, box annotations,
|
| 147 |
-
categories, and referring expressions) follows the same non-commercial
|
| 148 |
-
terms (`cc-by-nc-sa-4.0`). Redistribution of the underlying nuScenes images
|
| 149 |
-
is subject to nuScenes' own terms — see https://www.nuscenes.org/terms-of-use.
|
| 150 |
-
|
| 151 |
-
## Citation
|
| 152 |
-
|
| 153 |
-
If you use this dataset, please cite:
|
| 154 |
-
|
| 155 |
-
```
|
| 156 |
-
@inproceedings{chowdhury2026readtheobject,
|
| 157 |
-
title={Read the Object: A Dataset and Task for Text-Grounded Object Understanding in Real-World Images},
|
| 158 |
-
author={Chowdhury, Arpita and Park, Hyojin and Le, Hoang and Chao, Wei-Lun and Hayat, Munawar and Porikli, Fatih and Mahajan, Shweta},
|
| 159 |
-
}
|
| 160 |
-
```
|
| 161 |
-
|
| 162 |
-
and, for the underlying imagery, nuScenes:
|
| 163 |
-
|
| 164 |
-
```
|
| 165 |
-
@inproceedings{nuscenes2019,
|
| 166 |
-
title={nuScenes: A multimodal dataset for autonomous driving},
|
| 167 |
-
author={Caesar, Holger and Bankiti, Varun and Lang, Alex H. and others},
|
| 168 |
-
booktitle={CVPR},
|
| 169 |
-
year={2020}
|
| 170 |
-
}
|
| 171 |
-
```
|
|
|
|
| 77 |
| `num_texts_used` | int | how many text instances were combined into this sample |
|
| 78 |
| `sample_type` | string | `single_text_single_object`, `multi_text_single_object`, or `disambiguated_same_category` |
|
| 79 |
| `attribute` | string (optional) | present only for `disambiguated_same_category` samples — a distinguishing visual/spatial attribute added because another object of the *same category* in the same image shares the same text (e.g. `"left blue"` in `"'23' written on left blue gate"`) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|