Add DOTA dataset #2551
Add DOTA dataset #2551
Conversation
github-actions
bot
added
documentation
| @@ -14,6 +14,7 @@ Dataset,Task,Source,License,# Samples,# Classes,Size (px),Resolution (m),Bands | |||
| `DeepGlobe Land Cover`_,S,DigitalGlobe +Vivid,-,803,7,"2,448x2,448",0.5,RGB | |||
| `DFC2022`_,S,Aerial,"CC-BY-4.0","3,981",15,"2,000x2,000",0.5,RGB | |||
| `Digital Typhoon`_,"C, R",Himawari,"CC-BY-4.0","189,364",8,512,5000,Infrared | |||
| `DOTA`_,OD,"Google Earth, Gaofen-2, Jilin-1","CC-BY-NC-4.0","5,229",15,"varying","varying",RGB | |||
There was a problem hiding this comment.
I usually use a range of numbers instead of "varying"
There was a problem hiding this comment.
Where do you see this license? I agree that it's non-commercial, but there are multiple licenses like that.
There was a problem hiding this comment.
DOTA also includes aerial imagery provided by CycloMedia B.V.
The Google Earth images must adhere to their guidelines.
For other image sources, I did not mention a specific license in the dataset website, but it should be non-commercial.
| class DOTA(NonGeoDataset): | ||
| """DOTA dataset. | ||
|
|
||
| The `DOTA <https://captain-whu.github.io/DOTA/index.html>`_ is a large-scale object |
There was a problem hiding this comment.
| The `DOTA <https://captain-whu.github.io/DOTA/index.html>`_ is a large-scale object | |
| `DOTA <https://captain-whu.github.io/DOTA/index.html>`__ is a large-scale object |
| * helipad | ||
|
|
||
|
|
||
| If you use this work in your research, please cite the following paper: |
There was a problem hiding this comment.
| If you use this work in your research, please cite the following paper: | |
| If you use this work in your research, please cite the following papers: |
| Args: | ||
| root: root directory where dataset can be found | ||
| split: split of the dataset to use, one of ['train', 'val'] | ||
| version: version of the dataset to use, one of ['1.0', '2.0'] |
There was a problem hiding this comment.
Any reason you skipped 1.5? Is it very different?
There was a problem hiding this comment.
I suggest to also include DOTA-v1.5. The images of DOTA-v1.0 and DOTA-v1.5 are the same. But many small objects are annotated in DOTA-v1.5.
There was a problem hiding this comment.
Just a lot of work to format the data to common structure.
| split: str = 'train', | ||
| version: str = '2.0', | ||
| bbox_orientation: str = 'hbb', |
There was a problem hiding this comment.
| split: str = 'train', | |
| version: str = '2.0', | |
| bbox_orientation: str = 'hbb', | |
| split: Literal['train', 'val'] = 'train', | |
| version: Literal['1.0', '2.0'] = '2.0', | |
| bbox_orientation: Literal['hbb', 'obb'] = 'hbb', |
If you want to get technical, these type hints are more correct, but we don't yet use them anywhere else in TorchGeo (although we should).
| gsd:0.146343590398 | ||
|
|
||
| Args: | ||
| path: path to annotation file |
There was a problem hiding this comment.
Needs empty line after it
| split: split of the dataset to use, one of ['train', 'val'] | ||
| version: version of the dataset to use, one of ['1.0', '2.0'] | ||
| bbox_orientation: bounding box orientation, one of ['hbb', 'obb'], meaning horizontal | ||
| or oriented bounding boxes, hbb only available for v2.0 |
There was a problem hiding this comment.
I wonder if we could make our own hbb for 1.0 just by using min/max. It wouldn't be perfect, but it would be better than the current inconsistency?
There was a problem hiding this comment.
Hi, we have provided the hbb labels DOTA-v1.0.
There was a problem hiding this comment.
I haven't found any hbb labels for DOTA-v1.0 in the various folders across the google drive. For Version 2, the hbb labels are also just extracted from the obb labels.
There was a problem hiding this comment.
https://drive.google.com/file/d/1sS9hveKtYAiTsGVxC4msF5qJjhn3wYpY/view?usp=drive_link
@nilsleh The task 2 gt is for hbb labels
| if self.bbox_orientation == 'obb': | ||
| boxes, labels = self._load_annotations(sample_row['annotation_path']) | ||
| else: | ||
| boxes, labels = self._load_annotations(sample_row['annotation_path']) |
There was a problem hiding this comment.
Am I blind, or are these two lines the same?
| else: | ||
| boxes, labels = self._load_annotations(sample_row['annotation_path']) | ||
|
|
||
| sample['boxes'] = boxes |
There was a problem hiding this comment.
We no longer uses boxes for compatibility with Kornia, see #1978
This PR adds the DOTA dataset for object detection.
Dataset rehosted on HF for faster and comprehensive download in one place.
Dataset features:
Dataset format:
Horizontal BBox example:
Oriented BBox example:
@dingjiansw101 as inquired in CAPTAIN-WHU/DOTA#29, this is the PR that aims to make the DOTA datset more accessible. The dataset structure in the google and baidu drives was a bit confusing and I restructured it on HF which is hopefully more clear. If you would take a look and have any comments, that would be helpful.
Also @robmarkcole perhaps interesting for you.
Closes #1