Datasets:

JakeTurner616
/

mtg-cards-SIFT-Features

@@ -17,113 +17,125 @@ pretty_name: mtg card SIFT Feature dataset
 size_categories:
 - 100K<n<1M
 tags:
-- v5.0
 ---
-# MTG Card SIFT Features Dataset (v5)
 ![demo](https://github.com/JakeTurner616/simple-mtg-feature-extraction/raw/main/docs/feature-detect.gif)
-This dataset contains the latest output of my Magic: The Gathering (MTG) card image feature extraction pipeline, now upgraded to **version 5**. This release is designed for **incremental model growth**, allowing descriptor and ID map updates **without retraining from scratch** — ideal for server-side production use.
-> **Note:** Version 5 introduces additive updates to the FAISS index and `id_map.json`, enabling true incremental inference pipelines.
 ---
-## What’s New in v5?
-| Feature              | v4 (Previous)          | v5 (Current)                                   |
-| -------------------- | ---------------------- | ---------------------------------------------- |
-| Index update         | Full rebuild required  | Additive-safe incremental indexing             |
-| ID map behavior      | Rewritten each time    | Appended in-place to preserve alignment        |
-| Feature pipeline     | Batch-oriented         | Fully resumable and append-safe                |
-| Format compatibility | `.h5 + .json + .index` | Same, now stable for long-term use             |
-| RAM usage            | Low (<500MB typical)   | Still low, with improved batching and flushing |
 ---
-## File Layout (resourcesV5.zip)
 ```
 resources/
 └── run/
-    ├── candidate_features.h5     # HDF5 with keypoints + descriptors per card
-    ├── faiss_ivf.index           # FAISS IVF-PQ index (incremental-safe)
-    └── id_map.json               # Linear descriptor-to-scryfall_id mapping
 ```
 ---
 ## What Is This?
-This is a high-precision visual descriptor dataset for Magic: The Gathering cards, intended for computer vision applications such as visual search, real-world scanning, and inference pipelines.
-### Pipeline Summary
 1. **Image Acquisition**
-   * Cards are fetched via the Scryfall API.
 2. **Preprocessing**
-   * Images are resized with aspect ratio preserved, processed via CLAHE (L-channel), and converted to grayscale.
-3. **SIFT + RootSIFT**
-   * Keypoints are detected using OpenCV’s SIFT; RootSIFT normalization applied for FAISS compatibility.
-4. **Efficient Storage**
-   * All features and keypoints are stored in an HDF5 file (`candidate_features.h5`) with gzip compression.
 5. **Indexing**
-   * Descriptors are indexed using FAISS IVF-PQ or FlatL2 depending on scale.
-   * Index grows over time as new cards are processed.
 6. **Mapping**
-   * A flat `id_map.json` file aligns descriptor index rows to `scryfall_id`s.
 ---
-## Supported Use Cases
-* 🔍 **Visual Search**: Find the closest matching MTG card from a photo.
-* 🤖 **Model Inference**: Use as a lookup backend in a card recognition app.
-* 🔁 **Incremental Updates**: Continue indexing new cards without retraining.
-* 📦 **Dataset Analysis**: Explore descriptor density and clustering across cards.
 ---
-## Integration Workflow (Back-End)
 ```text
-[Input: Cropped Card Image]
-        ↓
-[CLAHE → Grayscale → SIFT → RootSIFT]
-        ↓
 [FAISS IVF-PQ Search]
-        ↓
-[Descriptor match → ID mapping]
-        ↓
-[Return most common scryfall_id]
 ```
 ---
-## Key Benefits
-✅ No SQLite or database state required
-✅ Works seamlessly with FAISS, NumPy, and h5py
-✅ Handles partial runs and resumes automatically
-✅ Small memory footprint and parallel-safe
-✅ MIT Licensed — freely usable and extendable
 ---
 ## Acknowledgments
-* Created by [JakeTurner616](https://github.com/JakeTurner616)
-* Based on images and data from [Scryfall](https://scryfall.com/docs/api)
-* Uses [FAISS](https://github.com/facebookresearch/faiss), [OpenCV](https://opencv.org/), and [h5py](https://www.h5py.org/)
 ---
-For more technical details, visit the [workflow notebook](https://github.com/JakeTurner616/simple-mtg-feature-extraction/blob/production/workflow.ipynb) or open an issue on GitHub.

 size_categories:
 - 100K<n<1M
 tags:
+- v5.1
 ---
+# MTG Card SIFT Features Dataset (v5.1)
 ![demo](https://github.com/JakeTurner616/simple-mtg-feature-extraction/raw/main/docs/feature-detect.gif)
+This dataset contains the **latest incremental MTG card SIFT + RootSIFT feature extraction pipeline**. It is designed for **server-side production inference**, enabling **additive updates to the FAISS index and `id_map.json`** without retraining or reindexing from scratch.
+> **Note:** This version aligns with a daily `resources-nightly.zip` Hugging Face upload workflow for reliable continuous deployment via my production server.
 ---
+## What’s New in v5.1?
+| Feature              | v5.0                       | v5.1 (Current)                      |
+| -------------------- | -------------------------- | ----------------------------------- |
+| Index updates        | Additive-safe              | Same                                |
+| Upload workflow      | Manual or ad-hoc           | Integrated with HF nightly pipeline |
+| Logging              | Basic                      | Detailed zipping and upload logs    |
+| Service impact       | Potential blocking uploads | Runs in background, non-blocking    |
+| HF dataset structure | Single zip upload          | Same, consistent naming             |
 ---
+## File Layout (`resources-nightly.zip`)
 ```
 resources/
 └── run/
+    ├── candidate_features.h5     # Keypoints + descriptors per card (gzip HDF5)
+    ├── faiss_ivf.index           # FAISS IVF-PQ index
+    └── id_map.json               # Descriptor index-to-scryfall_id mapping
 ```
 ---
 ## What Is This?
+A **high-precision visual descriptor dataset** for Magic: The Gathering cards, used for:
+✅ **Visual Search**
+✅ **Mobile/Server Card Scanning**
+✅ **FAISS-based Similarity Search**
+✅ **Incremental Model Growth**
+---
+## Pipeline Summary
 1. **Image Acquisition**
+   * Fetched from Scryfall using official API.
 2. **Preprocessing**
+   * Resize with aspect ratio, CLAHE on L-channel, grayscale conversion.
+3. **Feature Extraction**
+   * OpenCV SIFT + RootSIFT normalization.
+4. **Storage**
+   * `candidate_features.h5` using gzip compression.
 5. **Indexing**
+   * FAISS IVF-PQ (100 clusters, 8x8 PQ), additive updates supported.
 6. **Mapping**
+   * `id_map.json` aligns descriptors to `scryfall_id`.
+7. **HF Upload**
+   * Zipped nightly, uploaded to HF dataset repo in the background.
 ---
+## Use Cases
+✅ **Real-time card scanning** for Magic: The Gathering
+✅ **Card image search pipelines**
+✅ **Local inference on low-resource servers**
+✅ **Model growth without reindexing**
 ---
+## Workflow (Simplified)
 ```text
+[Input Image]
+    ↓
+[CLAHE + SIFT + RootSIFT]
+    ↓
 [FAISS IVF-PQ Search]
+    ↓
+[Retrieve scryfall_id]
+    ↓
+[Result]
 ```
 ---
+## Why v5.1 Matters
+✅ **No SQLite required**
+✅ **Works with h5py + NumPy**
+✅ **Parallel-safe and resumable**
+✅ **Tiny memory footprint**
+✅ **MIT Licensed for free, unlimited use**
 ---
 ## Acknowledgments
+Created by [JakeTurner616](https://github.com/JakeTurner616)
+Powered by:
+* [Scryfall API](https://scryfall.com/docs/api)
+* [FAISS](https://github.com/facebookresearch/faiss)
+* [OpenCV](https://opencv.org/)
+* [h5py](https://www.h5py.org/)
 ---
+For deep implementation details, see the [mtgscan.cards monorepo](https://github.com/mtgscan-cards/monolithic).