yeeecheng commited on
Commit
38754b5
·
verified ·
1 Parent(s): e82b380

Add dataset card

Browse files
Files changed (1) hide show
  1. README.md +97 -0
README.md ADDED
@@ -0,0 +1,97 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-sa-4.0
3
+ task_categories:
4
+ - image-to-image
5
+ - image-classification
6
+ tags:
7
+ - image-restoration
8
+ - super-resolution
9
+ - blind-image-super-resolution
10
+ - clip
11
+ - degradation
12
+ - lpips
13
+ size_categories:
14
+ - 100K<n<1M
15
+ pretty_name: OD-CLIP Training Dataset
16
+ ---
17
+
18
+ # OD-CLIP Training Dataset
19
+
20
+ Training data for **OD-CLIP**: Learning Ordinal Degradation Representations with
21
+ Textual Priors for Diffusion-Based Blind Image Super-Resolution
22
+ (AAAI'27 submission).
23
+
24
+ This dataset provides degraded low-quality (LQ) crops paired with per-image
25
+ degradation metadata, covering four synthetic degradation types (Gaussian blur,
26
+ Gaussian noise, JPEG compression, and downsampling) at a dense grid of physical
27
+ severity levels. It is used to train the OD-CLIP severity-aware image encoder.
28
+
29
+ ## Contents
30
+
31
+ ### 1. `DIV2K_HR_train_crop/`
32
+ Single-degradation training set. Contains 208 subdirectories, one per
33
+ `(degradation_type, level)` pair:
34
+
35
+ | Type | Physical parameter | Range | Step | # subdirs |
36
+ |:-----------|:-------------------|:---------------|:------|----------:|
37
+ | Blur | Gaussian sigma | 0.1 → 4.0 | 0.1 | 40 |
38
+ | JPEG | Quality factor | 31 → 95 | 1 | 65 |
39
+ | Downsample | Scale factor | 1.1 → 7.0 | 0.1 | 60 |
40
+ | Noise | Gaussian sigma | 1.0 → 40.0 | 1.0 | 40 |
41
+ | | | **Total** | | **205** |
42
+
43
+ *(Additional 3 clean-reference subdirs bring the total to 208.)*
44
+
45
+ Each subdirectory contains:
46
+ - `LQ/` — 800 degraded crops (256×256 px, PNG)
47
+ - `daclip_val.csv` — per-image manifest (filepath, degradation label, physical
48
+ parameter value)
49
+ - `degraded_prompts_0.json` — text prompt used for the degradation
50
+
51
+ ### 2. `DIV2K_train_HR_random/`
52
+ Mixed-degradation training set. All four degradations are applied to each
53
+ image with per-image randomized parameters, used for stage-2 fine-tuning on
54
+ composite degradations.
55
+
56
+ - `train_crop/random0/LQ/` — 4000 mixed-degradation crops
57
+ - `merged_daclip_train_all_updated.csv` — top-level manifest for stage-2 training
58
+
59
+ ## Note on Ground Truth
60
+
61
+ **Ground-truth (GT) crops are not included in this release** to reduce
62
+ distribution size. They can be regenerated from the publicly-available
63
+ DIV2K training set:
64
+ - Source: https://data.vision.ee.ethz.ch/cvl/DIV2K/
65
+ - Use the crop-and-degrade pipeline in the accompanying code repository.
66
+
67
+ ## Severity Calibration
68
+
69
+ Physical parameters `l` are mapped to a perceptual severity coordinate
70
+ `β_l^(t)` via LPIPS-based calibration. The mapping table used for training
71
+ is included in the accompanying code repository as
72
+ `lpips_guided_mapping.json`.
73
+
74
+ Details on the calibration protocol (LPIPS backbone, sample size, per-type
75
+ monotonicity properties) are provided in Appendix A of the paper.
76
+
77
+ ## License
78
+
79
+ Derived from DIV2K, licensed under **CC BY-SA 4.0**.
80
+ This release inherits the same license.
81
+
82
+ ## Citation
83
+
84
+ ```bibtex
85
+ @inproceedings{odclip2027,
86
+ title = {Learning Ordinal Degradation Representations with Textual Priors
87
+ for Diffusion-Based Blind Image Super-Resolution},
88
+ author = {Anonymous},
89
+ booktitle = {AAAI},
90
+ year = {2027}
91
+ }
92
+ ```
93
+
94
+ ## Contact
95
+
96
+ For questions or requests for the full pipeline (including GT + degradation
97
+ scripts), please open an issue on the associated code repository.