Upload README.md with huggingface_hub

README.md

@@ -4,12 +4,13 @@ task_categories:
   - object-detection
 tags:
   - stamp-detection
-  - yolo
   - document-analysis
+  - object-detection
+  - yolo
   - computer-vision
-pretty_name: Stamp Detection Dataset
+pretty_name: Clean Core Stamp Detection Dataset
 size_categories:
-  - 10K<n<100K
+  - 1K<n<10K
 dataset_info:
   features:
     - name: image
@@ -18,96 +19,91 @@ dataset_info:
       dtype: string
   splits:
     - name: train
-      num_examples: 20857
+      num_examples: 6283
     - name: validation
-      num_examples: 3059
+      num_examples: 785
     - name: test
-      num_examples: 3059
+      num_examples: 786
 ---
 
-# Stamp Detection Dataset
+# Clean Core Stamp Detection Dataset
 
-A large-scale, unified dataset for detecting stamps on scanned documents using YOLO-based object detection models.
+A cleaned document-stamp detection dataset built from curated raw sources only.
+This version replaces the previous `mapo80/stamps` release and removes noisy or
+off-target sources such as postage stamps, shape/identity datasets, corrupted
+multi-class remaps, and duplicated negatives.
 
 ## Overview
 
 | Parameter | Value |
 |-----------|-------|
-| **Task** | Object Detection |
-| **Classes** | 1 — `stamp` (class id 0) |
-| **Total images** | **26,975** |
-| **Positive images** | 21,037 (with stamp annotations) |
-| **Negative images** | 5,938 (background, no stamps) |
-| **Total bounding boxes** | 35,777 |
-| **Annotation format** | YOLO txt (`class x_center y_center width height`, normalized 0–1) |
-| **Dataset size** | ~6 GB |
-
-## Download
-
-The dataset is distributed as **zip archives** (one per split):
-
-```bash
-# Download all splits
-wget https://huggingface.co/datasets/mapo80/stamps/resolve/main/train.zip
-wget https://huggingface.co/datasets/mapo80/stamps/resolve/main/val.zip
-wget https://huggingface.co/datasets/mapo80/stamps/resolve/main/test.zip
-
-# Unzip into a single directory
-mkdir stamps-dataset && cd stamps-dataset
-unzip ../train.zip
-unzip ../val.zip
-unzip ../test.zip
-
-# Download config
-wget https://huggingface.co/datasets/mapo80/stamps/resolve/main/dataset.yaml
-```
+| Task | Object detection |
+| Classes | 1 (`stamp`, class id 0) |
+| Total images | 7854 |
+| Positive images | 6608 |
+| Negative images | 1246 |
+| Total bounding boxes | 12659 |
+| Annotation format | YOLO txt (`class x_center y_center width height`) |
 
 ## Splits
 
 | Split | Total Images | Positive | Negative | Bounding Boxes |
-|-------|-------------|----------|----------|---------------|
-| Train | 20,857 | 16,503 | 4,354 | 28,101 |
-| Val | 3,059 | 2,268 | 791 | 3,851 |
-| Test | 3,059 | 2,266 | 793 | 3,825 |
-| **Total** | **26,975** | **21,037** | **5,938** | **35,777** |
+|-------|-------------:|---------:|---------:|---------------:|
+| Train | 6283 | 5287 | 996 | 10046 |
+| Val | 785 | 660 | 125 | 1298 |
+| Test | 786 | 661 | 125 | 1315 |
 
-Negative (background) images are documents **without stamps** — they have empty `.txt` label files. These help reduce false positives during training. The negative ratio is ~22%, sourced from RVL-CDIP, Tobacco3482, and FUNSD.
+## What Changed
 
-## Directory Structure
+- Kept only in-scope document stamp sources
+- Filtered `stamp_detection_stampa` to the raw `stamp` class only
+- Removed postage-stamp, shape, identity, and corrupted mixed-class sources
+- Rebuilt negatives from `RVL-CDIP` and `FUNSD`
+- Dropped severe negative outliers (blank pages, almost-black pages, heavily degraded scans)
+- Rebuilt train/val/test from scratch from `data/raw`
 
-After unzipping:
+## Included Sources
 
-```
-stamps-dataset/
-├── images/
-│   ├── train/     # 20,857 images
-│   ├── val/       #  3,059 images
-│   └── test/      #  3,059 images
-├── labels/
-│   ├── train/     # 20,857 label files (.txt)
-│   ├── val/       #  3,059 label files
-│   └── test/      #  3,059 label files
-└── dataset.yaml   # YOLO config file
-```
+### Positive sources
 
-## Annotation Format
+- `stamp_detection_jsam`: 4385 images
+- `yolo_stamp_classify`: 1397 images
+- `stamp_detectation_marcos`: 365 images
+- `stamp_detection_stampa` (filtered): 461 images
 
-Standard YOLO format — one `.txt` file per image:
+### Negative sources
 
-```
-0 0.5234 0.4123 0.1456 0.1987
-0 0.7812 0.6543 0.0874 0.1125
-```
+- `neg_rvl_cdip`: 1048 images
+- `neg_funsd`: 198 images
+
+## Excluded Source Families
+
+- `detect_postage_stamp`
+- `stamp_class`
+- `stamp_shape`
+- `stamp_individual`
+- `stamp_recognition`
+- `stamp_detection_shujing`
+- `stamp_warisara`
+- `staver_yolo`
+- `neg_tobacco3482`
+
+## Directory Structure
 
-Each line: `class_id x_center y_center width height` (all values normalized 0–1). Empty files = negative images (no stamps present).
+The repository stores the dataset as zip archives:
 
-## Usage with YOLO
+```text
+train.zip
+val.zip
+test.zip
+dataset.yaml
+```
 
-```python
-from ultralytics import YOLO
+Each zip contains:
 
-model = YOLO("yolov9s.pt")
-results = model.train(data="dataset.yaml", imgsz=320, epochs=300)
+```text
+images/<split>/
+labels/<split>/
 ```
 
 ## `dataset.yaml`
@@ -123,89 +119,8 @@ names:
   - stamp
 ```
 
-## Dataset Creation Pipeline
-
-1. **Download** from Roboflow Universe (API), Kaggle (API)
-2. **Convert** heterogeneous formats (binary masks, multi-class labels) → unified YOLO single-class
-3. **Merge** with stratified 80/10/10 split by source dataset
-4. **Deduplicate** using perceptual hashing (phash 16-bit) — removed 2,112 duplicates
-5. **Validate** bounding boxes (min area ≥ 0.1%, aspect ratio ≤ 10, within image bounds) — removed 393 invalid entries
-6. **Add negatives** from RVL-CDIP, Tobacco3482, FUNSD document datasets
-
-## Source Datasets
-
-### Roboflow Universe (CC BY 4.0)
-
-| # | Dataset | Author | Used Images | Link |
-|---|---------|--------|------------|------|
-| 1 | Stamp Detection | Stampa | 1,271 | [universe.roboflow.com/stampa/stamp-detection-4jhgg](https://universe.roboflow.com/stampa/stamp-detection-4jhgg) |
-| 2 | Stamp Detection | JSam | 4,539 | [universe.roboflow.com/jsam/stamp-detection-okgih](https://universe.roboflow.com/jsam/stamp-detection-okgih) |
-| 3 | Stamp Recognition | Stamp Project | 4,488 | [universe.roboflow.com/stamp-project/stamp-recognition-sdoan](https://universe.roboflow.com/stamp-project/stamp-recognition-sdoan/dataset/2) |
-| 4 | Stamp | class | 2,289 | [universe.roboflow.com/class-cemy0/stamp-wnh42](https://universe.roboflow.com/class-cemy0/stamp-wnh42) |
-| 5 | Stamp Detection | shujing liang | 2,341 | [universe.roboflow.com/shujing-liang/stamp-detection-f3yka](https://universe.roboflow.com/shujing-liang/stamp-detection-f3yka) |
-| 6 | YOLO-Stamp | Detect and Classify | 1,400 | [universe.roboflow.com/detect-and-classify/yolo-stamp-6jf5w](https://universe.roboflow.com/detect-and-classify/yolo-stamp-6jf5w) |
-| 7 | stamp-individual | swp | 2,992 | [universe.roboflow.com/swp-3jks1/stamp-individual](https://universe.roboflow.com/swp-3jks1/stamp-individual) |
-| 8 | stamp-shape | swp | 3,000 | [universe.roboflow.com/swp-3jks1/stamp-shape](https://universe.roboflow.com/swp-3jks1/stamp-shape/dataset/4) |
-| 9 | Stamp Detectation | Marcos | 427 | [universe.roboflow.com/marcos-7aslt/stamp-detectation](https://universe.roboflow.com/marcos-7aslt/stamp-detectation) |
-| 10 | Detect Postage Stamp | jackwild | 395 | [universe.roboflow.com/jackwildgooglecom/detect-postage-stamp](https://universe.roboflow.com/jackwildgooglecom/detect-postage-stamp) |
-
-### Kaggle
-
-| # | Dataset | Author | Used Images | License | Link |
-|---|---------|--------|------------|---------|------|
-| 11 | StaVer (Stamp Verification) | Rachael Tatman / DFKI | 400 | Free access | [kaggle.com/datasets/rtatman/stamp-verification-staver-dataset](https://www.kaggle.com/datasets/rtatman/stamp-verification-staver-dataset) |
-
-### Negative/Background Sources
-
-| # | Dataset | Used Images | License | Link |
-|---|---------|------------|---------|------|
-| 12 | RVL-CDIP | 1,060 | Free for research | [huggingface.co/datasets/chainyo/rvl-cdip](https://huggingface.co/datasets/chainyo/rvl-cdip) |
-| 13 | Tobacco3482 | 11 | Free | [kaggle.com/datasets/patrickaudriaz/tobacco3482jpg](https://www.kaggle.com/datasets/patrickaudriaz/tobacco3482jpg) |
-| 14 | FUNSD | 199 | Free for research | [huggingface.co/datasets/nielsr/funsd](https://huggingface.co/datasets/nielsr/funsd) |
-
-## Licenses
-
-The majority of this dataset is licensed under **CC BY 4.0** (Roboflow Universe sources). Other sources are freely available for research.
-
-| Source | License |
-|--------|---------|
-| Roboflow Universe (10 datasets) | **CC BY 4.0** |
-| StaVer (Kaggle) | Free access |
-| RVL-CDIP | Free for research |
-| Tobacco3482 | Free |
-| FUNSD | Free for research |
-
-### Attribution (CC BY 4.0)
-
-This dataset uses data from the following Roboflow Universe authors:
-- **Stampa** — Stamp Detection
-- **JSam** — Stamp Detection
-- **Stamp Project** — Stamp Recognition
-- **class** — Stamp Dataset
-- **shujing liang** — Stamp Detection
-- **DETECT AND CLASSIFY** — YOLO-Stamp
-- **swp** — stamp-individual, stamp-shape
-- **Marcos** — Stamp Detectation
-- **jackwildgooglecom** — Detect Postage Stamp
-
-## Citations
-
-```bibtex
-@inproceedings{micenkova2011stamp,
-  title={Stamp detection in color document images},
-  author={Micenkova, Barbora and van Beusekom, Joost},
-  booktitle={International Conference on Document Analysis and Recognition (ICDAR)},
-  pages={1125--1129},
-  year={2011},
-  organization={IEEE}
-}
-
-@inproceedings{harley2015rvlcdip,
-  title={Evaluation of Deep Convolutional Nets for Document Image Classification and Retrieval},
-  author={Harley, Adam W. and Ufkes, Alex and Derpanis, Konstantinos G.},
-  booktitle={International Conference on Document Analysis and Recognition (ICDAR)},
-  pages={991--995},
-  year={2015},
-  organization={IEEE}
-}
-```
+## Notes
+
+- Empty label files are valid negatives
+- This benchmark is intentionally narrower and cleaner than the previous release
+- It is optimized for document-level stamp detection, not postage stamps or stamp classification