Datasets:
Add dataset card
Browse files
README.md
CHANGED
|
@@ -23,7 +23,7 @@ task_categories:
|
|
| 23 |
|
| 24 |
## Why this dataset
|
| 25 |
|
| 26 |
-
Faces are one of the most perceptually sensitive regions in an image and there exist strong, well-studied priors for facial quality. Face2Scene leverages this: by treating the face as an oracle for degradation, the model learns to restore the surrounding scene. To support this, InScene provides images in which a consistent set of **identities** appear across many diverse scenes, together with the localization metadata (face boxes and landmarks) needed to isolate and reason about the facial region.
|
| 27 |
|
| 28 |
## Data generation
|
| 29 |
|
|
@@ -35,34 +35,44 @@ Please refer to that repository for the full generation details (identity condit
|
|
| 35 |
|
| 36 |
## Dataset structure
|
| 37 |
|
| 38 |
-
The dataset is
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
|
| 40 |
```
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
...
|
| 46 |
-
val/
|
| 47 |
-
...
|
| 48 |
```
|
| 49 |
|
| 50 |
-
|
|
|
|
| 51 |
|
| 52 |
### Splits
|
| 53 |
|
| 54 |
-
| Split |
|
| 55 |
-
|---------|--------:|------------------
|
| 56 |
-
| `train` |
|
| 57 |
-
| `val` |
|
| 58 |
|
| 59 |
-
*(The `train` and `val` identity sets are disjoint.)*
|
| 60 |
|
| 61 |
### Per-sample JSON schema
|
| 62 |
|
|
|
|
|
|
|
| 63 |
| Field | Type | Description |
|
| 64 |
|--------------|-----------------|-----------------------------------------------------------------------------|
|
| 65 |
-
| `image_name` | string | Filename of the paired PNG.
|
| 66 |
| `prompt` | string | Structured scene prompt used for generation (subject, framing, background, detail/style). |
|
| 67 |
| `image_size` | object | `{ "width": 1024, "height": 1024 }`. |
|
| 68 |
| `face_bbox` | list[int] | Face bounding box `[x1, y1, x2, y2]` in pixel coordinates. |
|
|
@@ -79,7 +89,7 @@ import glob, json, os
|
|
| 79 |
from PIL import Image
|
| 80 |
from huggingface_hub import snapshot_download
|
| 81 |
|
| 82 |
-
# Download
|
| 83 |
local = snapshot_download(
|
| 84 |
"amir477/InScene",
|
| 85 |
repo_type="dataset",
|
|
@@ -90,7 +100,7 @@ samples = sorted(glob.glob(os.path.join(local, "train", "*")))
|
|
| 90 |
folder = samples[0]
|
| 91 |
name = os.path.basename(folder)
|
| 92 |
|
| 93 |
-
image = Image.open(os.path.join(folder, name + ".png")) # 1024x1024 RGB
|
| 94 |
meta = json.load(open(os.path.join(folder, name + ".json")))
|
| 95 |
|
| 96 |
print(image.size) # (1024, 1024)
|
|
@@ -99,6 +109,15 @@ print(meta["face_bbox"]) # [x1, y1, x2, y2]
|
|
| 99 |
print(len(meta["landmark"])) # 68
|
| 100 |
```
|
| 101 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
## Citation
|
| 103 |
|
| 104 |
If you find these images useful, please cite:
|
|
|
|
| 23 |
|
| 24 |
## Why this dataset
|
| 25 |
|
| 26 |
+
Faces are one of the most perceptually sensitive regions in an image and there exist strong, well-studied priors for facial quality. Face2Scene leverages this: by treating the face as an oracle for degradation, the model learns to restore the surrounding scene. To support this, InScene provides images in which a consistent set of **identities** appear across many diverse scenes, together with the localization metadata (face boxes and landmarks) needed to isolate and reason about the facial region. The `val` split additionally ships fixed **low-quality (degraded) inputs** paired with their clean targets, for consistent evaluation of restoration.
|
| 27 |
|
| 28 |
## Data generation
|
| 29 |
|
|
|
|
| 35 |
|
| 36 |
## Dataset structure
|
| 37 |
|
| 38 |
+
The dataset is split into `train/` and `val/`. Each split is a flat directory of **sample folders** (one folder per generated image). The folder name encodes provenance: `iden_<identity>_img_<source-face-id>_sample_<sample-id>_repeat_<k>` (the same prompt/identity may be sampled multiple times).
|
| 39 |
+
|
| 40 |
+
**`train`** — each folder holds the generated high-quality image and its metadata:
|
| 41 |
+
|
| 42 |
+
```
|
| 43 |
+
train/iden_00001_img_11_sample_321_repeat_1/
|
| 44 |
+
iden_00001_img_11_sample_321_repeat_1.png # 1024x1024 RGB image (HQ)
|
| 45 |
+
iden_00001_img_11_sample_321_repeat_1.json # metadata
|
| 46 |
+
```
|
| 47 |
+
|
| 48 |
+
**`val`** — each folder additionally holds a degraded, low-quality version of the image prefixed `LQ_`. The un-prefixed `*.png` is the **clean target** and the `LQ_*.png` is the **degraded restoration input**:
|
| 49 |
|
| 50 |
```
|
| 51 |
+
val/iden_00003_img_4_sample_473_repeat_1/
|
| 52 |
+
iden_00003_img_4_sample_473_repeat_1.png # 1024x1024 RGB image (HQ target)
|
| 53 |
+
LQ_iden_00003_img_4_sample_473_repeat_1.png # degraded input
|
| 54 |
+
iden_00003_img_4_sample_473_repeat_1.json # metadata
|
|
|
|
|
|
|
|
|
|
| 55 |
```
|
| 56 |
|
| 57 |
+
> Note: 6 of the val folders also contain a superseded copy of the trio prefixed `old_`
|
| 58 |
+
> (`old_*.png`, `old_LQ_*.png`, `old_*.json`), kept for provenance.
|
| 59 |
|
| 60 |
### Splits
|
| 61 |
|
| 62 |
+
| Split | Sample folders | Identities | Files | Size |
|
| 63 |
+
|---------|---------------:|-----------:|-------|------:|
|
| 64 |
+
| `train` | 11,260 | 870 | 11,260 HQ PNG + 11,260 JSON | 19 GB |
|
| 65 |
+
| `val` | 1,329 | 100 | 1,329 HQ PNG + 1,329 LQ PNG + 1,329 JSON (+ 18 `old_*` leftover files) | 3.4 GB |
|
| 66 |
|
| 67 |
+
*(The `train` and `val` identity sets are disjoint. `train` contains no LQ images — degradation for training is expected to be applied on the fly.)*
|
| 68 |
|
| 69 |
### Per-sample JSON schema
|
| 70 |
|
| 71 |
+
The JSON describes the **HQ** image (`image_name` is the un-prefixed PNG); for `val`, the degraded counterpart is the same name prefixed with `LQ_`.
|
| 72 |
+
|
| 73 |
| Field | Type | Description |
|
| 74 |
|--------------|-----------------|-----------------------------------------------------------------------------|
|
| 75 |
+
| `image_name` | string | Filename of the paired HQ PNG. |
|
| 76 |
| `prompt` | string | Structured scene prompt used for generation (subject, framing, background, detail/style). |
|
| 77 |
| `image_size` | object | `{ "width": 1024, "height": 1024 }`. |
|
| 78 |
| `face_bbox` | list[int] | Face bounding box `[x1, y1, x2, y2]` in pixel coordinates. |
|
|
|
|
| 89 |
from PIL import Image
|
| 90 |
from huggingface_hub import snapshot_download
|
| 91 |
|
| 92 |
+
# Download one split (use allow_patterns="val/*" for val)
|
| 93 |
local = snapshot_download(
|
| 94 |
"amir477/InScene",
|
| 95 |
repo_type="dataset",
|
|
|
|
| 100 |
folder = samples[0]
|
| 101 |
name = os.path.basename(folder)
|
| 102 |
|
| 103 |
+
image = Image.open(os.path.join(folder, name + ".png")) # 1024x1024 RGB (HQ)
|
| 104 |
meta = json.load(open(os.path.join(folder, name + ".json")))
|
| 105 |
|
| 106 |
print(image.size) # (1024, 1024)
|
|
|
|
| 109 |
print(len(meta["landmark"])) # 68
|
| 110 |
```
|
| 111 |
|
| 112 |
+
For the `val` split, each sample also has a degraded input alongside the clean target:
|
| 113 |
+
|
| 114 |
+
```python
|
| 115 |
+
val_folder = sorted(glob.glob(os.path.join(local_val, "val", "*")))[0]
|
| 116 |
+
name = os.path.basename(val_folder)
|
| 117 |
+
target = Image.open(os.path.join(val_folder, name + ".png")) # clean HQ target
|
| 118 |
+
degraded = Image.open(os.path.join(val_folder, "LQ_" + name + ".png")) # degraded input
|
| 119 |
+
```
|
| 120 |
+
|
| 121 |
## Citation
|
| 122 |
|
| 123 |
If you find these images useful, please cite:
|