Datasets:
Commit
•
90e2875
1
Parent(s):
9d3ff07
Update README.md
Browse files
README.md
CHANGED
@@ -66,10 +66,10 @@ This dataset contains a subset of data used in the paper [You Actually Look Twic
|
|
66 |
|
67 |
## Dataset Structure
|
68 |
|
69 |
-
This dataset has two configurations. These configurations both cover the same data and annotations but provide these annotations in different forms to make it easier to
|
70 |
|
71 |
-
- The first configuration `YOLO
|
72 |
-
- The second configuration converts the YOLO format into a format which is closer to the `COCO` annotation format. This is done
|
73 |
|
74 |
### Data Instances
|
75 |
|
@@ -200,21 +200,21 @@ An example instance from the YOLO config:
|
|
200 |
The fields for the YOLO config:
|
201 |
|
202 |
- `image`: the image
|
203 |
-
- `objects`: the annotations which
|
204 |
- `bbox`: a list of bounding boxes for the image
|
205 |
- `label`: a list of labels for this image
|
206 |
|
207 |
The fields for the COCO config:
|
208 |
|
209 |
-
- `
|
210 |
- `width`: width of the image
|
211 |
- `image`: image
|
212 |
- `image_id`: id for the image
|
213 |
- `objects`: annotations in COCO format, consisting of a list containing dictionaries with the following keys:
|
214 |
- `bbox`: bounding boxes for the images
|
215 |
-
- `category_id`: label for the image
|
216 |
- `image_id`: id for the image
|
217 |
-
- `iscrowd`: COCO
|
218 |
- `segmentation`: COCO segmentation annotations (empty in this case but kept for compatibility with other processing scripts)
|
219 |
|
220 |
|
@@ -231,11 +231,11 @@ The dataset contains a train, validation and test split with the following numbe
|
|
231 |
|
232 |
## Dataset Creation
|
233 |
|
234 |
-
> [this] dataset was produced using a single source, the Lectaurep Repertoires dataset [Rostaing et al., 2021], which served as a basis for only the training and development split. The testset is composed of original data, from various documents, from the 17th century up to the early 20th with a single soldier war report. The test set is voluntarily very different and out of
|
235 |
.
|
236 |
### Curation Rationale
|
237 |
|
238 |
-
This dataset was created to produce a simplified version of the [Lectaurep Repertoires dataset](https://github.com/HTR-United/lectaurep-repertoires) which was found to contain:
|
239 |
|
240 |
> around 16 different ways to describe columns, from Col1 to Col7, the case-different col1-col7 and finally ColPair and ColOdd, which we all reduced to Col p.8
|
241 |
|
@@ -281,11 +281,11 @@ This data does not contain information relating to living individuals.
|
|
281 |
|
282 |
### Social Impact of Dataset
|
283 |
|
284 |
-
|
285 |
|
286 |
### Discussion of Biases
|
287 |
|
288 |
-
Historical documents contain a wide variety of page layouts. This means that the ability of models trained on this dataset to transfer to documents
|
289 |
|
290 |
### Other Known Limitations
|
291 |
|
|
|
66 |
|
67 |
## Dataset Structure
|
68 |
|
69 |
+
This dataset has two configurations. These configurations both cover the same data and annotations but provide these annotations in different forms to make it easier to integrate the data with existing processing pipelines.
|
70 |
|
71 |
+
- The first configuration, `YOLO`, uses the data's original format.
|
72 |
+
- The second configuration converts the YOLO format into a format which is closer to the `COCO` annotation format. This is done to make it easier to work with the `feature_extractor`s from the `Transformers` models for object detection, which expect data to be in a COCO style format.
|
73 |
|
74 |
### Data Instances
|
75 |
|
|
|
200 |
The fields for the YOLO config:
|
201 |
|
202 |
- `image`: the image
|
203 |
+
- `objects`: the annotations which consist of:
|
204 |
- `bbox`: a list of bounding boxes for the image
|
205 |
- `label`: a list of labels for this image
|
206 |
|
207 |
The fields for the COCO config:
|
208 |
|
209 |
+
- `height`: height of the image
|
210 |
- `width`: width of the image
|
211 |
- `image`: image
|
212 |
- `image_id`: id for the image
|
213 |
- `objects`: annotations in COCO format, consisting of a list containing dictionaries with the following keys:
|
214 |
- `bbox`: bounding boxes for the images
|
215 |
+
- `category_id`: a label for the image
|
216 |
- `image_id`: id for the image
|
217 |
+
- `iscrowd`: COCO `iscrowd` flag
|
218 |
- `segmentation`: COCO segmentation annotations (empty in this case but kept for compatibility with other processing scripts)
|
219 |
|
220 |
|
|
|
231 |
|
232 |
## Dataset Creation
|
233 |
|
234 |
+
> [this] dataset was produced using a single source, the Lectaurep Repertoires dataset [Rostaing et al., 2021], which served as a basis for only the training and development split. The testset is composed of original data, from various documents, from the 17th century up to the early 20th with a single soldier war report. The test set is voluntarily very different and out of domain with column borders that are not drawn nor printed in certain cases, layout in some kind of masonry layout. p.8
|
235 |
.
|
236 |
### Curation Rationale
|
237 |
|
238 |
+
This dataset was created to produce a simplified version of the [Lectaurep Repertoires dataset](https://github.com/HTR-United/lectaurep-repertoires), which was found to contain:
|
239 |
|
240 |
> around 16 different ways to describe columns, from Col1 to Col7, the case-different col1-col7 and finally ColPair and ColOdd, which we all reduced to Col p.8
|
241 |
|
|
|
281 |
|
282 |
### Social Impact of Dataset
|
283 |
|
284 |
+
A growing number of datasets are related to page layout for historical documents. This dataset offers a different approach to annotating these datasets (focusing on object detection rather than pixel-level annotations). Improving document layout recognition can have a positive impact on downstream tasks, in particular Optical Character Recognition.
|
285 |
|
286 |
### Discussion of Biases
|
287 |
|
288 |
+
Historical documents contain a wide variety of page layouts. This means that the ability of models trained on this dataset to transfer to documents with very different layouts is not guaranteed.
|
289 |
|
290 |
### Other Known Limitations
|
291 |
|