HUG-ABC commited on
Commit
08230ec
·
verified ·
1 Parent(s): 3f7a063

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +35 -21
README.md CHANGED
@@ -16,25 +16,21 @@ tags:
16
 
17
  # Delta-QA
18
 
19
- ## Dataset Overview
20
 
21
- Delta-QA accompanies [*Decoding the Delta: Unifying Remote Sensing Change Detection and Understanding with Multimodal Large Language Models*](https://arxiv.org/abs/2604.14044), published at ACM Multimedia 2026. It supports question answering over multi-temporal remote-sensing imagery for change detection and change understanding.
22
 
23
  **Authors:** Xiaohe Li, Jiahao Li, Kaixin Zhang, Yuqiang Fang, Leilei Lin, Hong Wang, Haohua Wu, and Zide Fan.
24
 
25
- This release contains the dataset only. Delta-LLaVA model code and weights are not included. The archives preserve the source image and JSON annotation organization without rewriting annotation records.
26
-
27
  ## Dataset Composition
28
 
29
- Delta-QA contains three independently identified components, each retaining its original `train` and `test` split:
30
-
31
- - **SECOND** uses paired temporal images under `im1/` and `im2/`, together with JSON question-answer annotations.
32
- - **LandsatSCD** uses paired temporal images under `im1/` and `im2/`, together with JSON question-answer annotations.
33
- - **WUSU-SCD** uses three temporal image directories, `im1/`, `im2/`, and `im3/`, together with JSON question-answer annotations. Here, `im3/` is the third temporal image directory rather than an auxiliary label directory.
34
 
35
- The components are distributed as separate ZIP archives so that their dataset identities and source splits remain explicit.
 
 
36
 
37
- ## Dataset Structure
38
 
39
  ```text
40
  Delta-QA/
@@ -51,12 +47,26 @@ Delta-QA/
51
  └── test.zip
52
  ```
53
 
54
- Within each archive, the original image directories and JSON annotation paths are retained. Annotation records associate questions with image samples through fields such as `image_id` and `file_name`. Depending on the task, a record may contain a natural-language `question`, an `answer`, answer choices, `image_size`, a `visual_prompt`, and `groundings`.
 
 
 
 
55
 
56
- A visual prompt identifies an image-space cue, such as a point or bounding box, when one is supplied by the task. Grounding annotations connect answer entities to spatial masks and may include answer character spans and COCO compressed run-length encoding (RLE). Field availability varies across components and tasks. Coordinates and annotations are preserved as supplied, and users should not assume that every JSON file has an identical schema.
 
 
 
 
 
 
 
 
57
 
58
  ## Usage
59
 
 
 
60
  ```python
61
  from huggingface_hub import snapshot_download
62
 
@@ -66,7 +76,7 @@ dataset_dir = snapshot_download(
66
  )
67
  ```
68
 
69
- The raw annotations can be read directly from an archive without extracting the image collection:
70
 
71
  ```python
72
  import json
@@ -77,7 +87,7 @@ archive_path = Path(dataset_dir) / "data" / "SECOND" / "train.zip"
77
 
78
  with zipfile.ZipFile(archive_path) as archive:
79
  json_paths = sorted(
80
- name for name in archive.namelist() if name.lower().endswith(".json")
81
  )
82
  with archive.open(json_paths[0]) as stream:
83
  records = json.load(stream)
@@ -87,18 +97,22 @@ print(example["question"])
87
  print(example["answer"])
88
  ```
89
 
90
- Use `file_name` or the corresponding `image_id` to locate matching image files under `im1/` and `im2/`. WUSU-SCD additionally provides the corresponding third temporal image under `im3/`.
91
 
92
- ## Data Sources and License
93
 
94
- - **SECOND:** source project page: <https://captain-whu.github.io/SCD/>. The audited source page provides download and citation information, but no explicit redistribution license was identified. Permission to re-host the images and source-derived labels remains **to be confirmed before public release**.
95
- - **LandsatSCD:** source deposit: <https://doi.org/10.6084/m9.figshare.19946135>; source paper: <https://doi.org/10.1080/17538947.2022.2111470>. The Figshare deposit identifies the dataset as CC BY 4.0. Redistribution must retain the required attribution and license notice.
96
- - **WUSU-SCD:** source repository: <https://github.com/AngieNikki/openWUSU>; source paper: <https://doi.org/10.1080/17538947.2023.2246445>. The source repository mentions ODbL/DbCL terms together with an academic-only, non-commercial restriction. The relationship between those statements, and the redistribution status of the underlying GF-2 imagery, remains **to be confirmed in writing before public release**.
97
 
98
- No single license is assigned to the complete Delta-QA release at this stage. The license for Delta-QA-authored question-answer annotations and permissions governing source-derived groundings must be finalized after the component-level rights are resolved. Local preparation of these archives does not constitute permission for public redistribution.
 
 
 
 
99
 
100
  ## Citation
101
 
 
 
102
  ```bibtex
103
  @inproceedings{li2026decoding,
104
  author = {Li, Xiaohe and Li, Jiahao and Zhang, Kaixin and Fang, Yuqiang and Lin, Leilei and Wang, Hong and Wu, Haohua and Fan, Zide},
 
16
 
17
  # Delta-QA
18
 
19
+ Delta-QA is a visual question answering dataset for multi-temporal remote sensing change detection and understanding. Built upon SECOND, LandsatSCD, and WUSU-SCD, it contains question-answer annotations associated with bi-temporal and tri-temporal imagery. The annotations cover image-level change understanding and spatially prompted questions, with grounding information for applicable tasks.
20
 
21
+ **Paper:** [Decoding the Delta: Unifying Remote Sensing Change Detection and Understanding with Multimodal Large Language Models](https://arxiv.org/abs/2604.14044) (ACM Multimedia 2026)
22
 
23
  **Authors:** Xiaohe Li, Jiahao Li, Kaixin Zhang, Yuqiang Fang, Leilei Lin, Hong Wang, Haohua Wu, and Zide Fan.
24
 
 
 
25
  ## Dataset Composition
26
 
27
+ Delta-QA contains three components, each organized into training and test archives:
 
 
 
 
28
 
29
+ - **SECOND:** Bi-temporal imagery stored in `im1/` and `im2/`, with corresponding question-answer annotations.
30
+ - **LandsatSCD:** Bi-temporal imagery stored in `im1/` and `im2/`, with corresponding question-answer annotations.
31
+ - **WUSU-SCD:** Tri-temporal imagery stored in `im1/`, `im2/`, and `im3/`, with corresponding question-answer annotations.
32
 
33
+ ## File Organization
34
 
35
  ```text
36
  Delta-QA/
 
47
  └── test.zip
48
  ```
49
 
50
+ Each archive contains image directories and JSON annotation files. Image references and annotation structures are preserved in the distributed archives.
51
+
52
+ ## Annotation Format
53
+
54
+ Annotation records associate questions with image samples through fields including:
55
 
56
+ | Field | Description |
57
+ | --- | --- |
58
+ | `image_id`, `file_name` | Image identifiers and file references. |
59
+ | `question`, `answer` | Question text and corresponding answer. |
60
+ | `image_size` | Spatial dimensions of the referenced image. |
61
+ | `visual_prompt` | Spatial prompts, including points or bounding boxes where applicable. |
62
+ | `groundings` | Region-level grounding information, including masks and answer-text spans where available. |
63
+
64
+ Answer options are included for applicable question types. Grounding masks may use COCO compressed run-length encoding (RLE). Field availability varies by task and dataset component.
65
 
66
  ## Usage
67
 
68
+ Download the dataset with `huggingface_hub`:
69
+
70
  ```python
71
  from huggingface_hub import snapshot_download
72
 
 
76
  )
77
  ```
78
 
79
+ Read JSON annotations directly from an archive:
80
 
81
  ```python
82
  import json
 
87
 
88
  with zipfile.ZipFile(archive_path) as archive:
89
  json_paths = sorted(
90
+ path for path in archive.namelist() if path.lower().endswith(".json")
91
  )
92
  with archive.open(json_paths[0]) as stream:
93
  records = json.load(stream)
 
97
  print(example["answer"])
98
  ```
99
 
100
+ Image samples are referenced by `image_id` and `file_name` in the annotation records.
101
 
102
+ ## Source Datasets and Licensing
103
 
104
+ Delta-QA is derived from the following remote sensing datasets:
 
 
105
 
106
+ - **SECOND:** [Official dataset website](https://captain-whu.github.io/SCD/) | [Source repository](https://github.com/CAPTAIN-WHU/SCD)
107
+ - **LandsatSCD:** [Dataset](https://doi.org/10.6084/m9.figshare.19946135) ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) | [Source publication](https://doi.org/10.1080/17538947.2022.2111470)
108
+ - **WUSU-SCD:** [Official dataset website](https://rsidea.whu.edu.cn/resource_wusu_sharing.htm) | [Source repository and usage terms](https://github.com/AngieNikki/openWUSU) | [Source publication](https://doi.org/10.1080/17538947.2023.2246445). The source project specifies academic, non-commercial use.
109
+
110
+ Source imagery and source-derived annotations remain subject to the applicable terms of their respective datasets. Please cite the original datasets when using their resources.
111
 
112
  ## Citation
113
 
114
+ If you use Delta-QA in your research, please cite:
115
+
116
  ```bibtex
117
  @inproceedings{li2026decoding,
118
  author = {Li, Xiaohe and Li, Jiahao and Zhang, Kaixin and Fang, Yuqiang and Lin, Leilei and Wang, Hong and Wu, Haohua and Fan, Zide},