thethreemagi/IYmtg
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs: Rewrite Step 3 - Damage Simulation Lab

Browse Source 
Critical fixes:
- Explained that Condition_Data uses Object Detection (bounding box
  annotations required in Create ML), not Image Classification
- Documented the grading formula from ConditionEngine: Inking/Rips/
  WaterDamage = Damaged; 0 = NM; 1-2 = EX; 3+ = PL
- Added materials list with all supplies needed
- Added card sourcing guide (bulk eBay lots, pre-damaged cards)
- Added raking light setup instructions
- Added folder name column to technique table (exact label names)
- Added clean card baseline requirement per damage type
- Expanded annotation guidance for Create ML

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Mike Wichers
2026-03-05 13:07:30 -05:00
parent b848a7b169
commit bb4ad9eb7e
1 changed files with 88 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

104
README.md
View File

@@ -388,23 +388,95 @@ Acquire pairs of cards to train the `StampDetector` (Promo/Date Stamped vs. Regu
* **Action:** Place cropped images of promos in `Stamp_Data/Stamped` and regular versions in `Stamp_Data/Clean`.
### Step 3: The "Damage Simulation Lab"
Techniques to ethically create training data using "Draft Chaff" (worthless cards).
| Category | Damage Type | Simulation Technique | Capture Tip (Crucial for ML) |
| :--- | :--- | :--- | :--- |
| **Surface** | Light Scratches | Rub foil surface gently with 0000 Steel Wool. | Use flash or moving light source to catch glint. |
| **Surface** | Clouding | Rub white eraser vigorously over foil surface. | Compare side-by-side with a clean card. |
| **Surface** | Dirt | Smudge lightly with potting soil or cocoa powder. | Ensure contrast against card art. |
| **Surface** | Dents | Press a ballpoint pen cap firmly into the surface. | **Raking Light:** Light from side to cast shadows. |
| **Edges** | Whitening | Rub card edges rapidly against denim jeans. | Photograph against a **Black Background**. |
| **Edges** | Chipping | Flake off small bits of black border. | Photograph against a **White Background**. |
| **Edges** | Corner Wear | Rub corners against a rough mousepad. | Macro focus on the corner radius. |
| **Structure** | Creases | Fold a corner until a hard line forms. | Catch the light reflection off the crease ridge. |
| **Structure** | Shuffle Bend | Riffle shuffle aggressively to create an arch. | Profile view (side view) to show curvature. |
| **Structure** | Water Damage | Mist with spray bottle, wait 60s, dry. | Catch the rippled surface texture with side light. |
| **Critical** | Inking | Use a Black Sharpie to "fix" whitened edges. | Use UV/Blacklight if possible, or bright white light. |
| **Critical** | Rips | Tear the edge slightly (approx. 5mm). | High contrast background. |
| **Critical** | Binders Dents | Press a 3-ring binder ring into the card. | Raking light to show the circular crimp. |
#### Important: This Model Uses Object Detection, Not Image Classification
The `Condition_Data` model is trained as an **Object Detection** model, not an Image Classification model. This is a critical difference:
- **Image Classification** (used for Foil and Stamp): drop images in a folder, Create ML labels them by folder name. Simple.
- **Object Detection** (used for Condition): you must **draw bounding boxes** around each defect in Create ML. The model learns *where* damage is on the card, not just that damage exists.
When training in Create ML, you will annotate each training image by drawing a rectangle around the damaged area and labeling it (e.g., "LightScratches", "Whitening"). Create ML has a built-in annotation tool — click an image, draw a box, type the label.
**Folder naming maps directly to label names.** The labels must match the `Condition_Data` subfolder names exactly:
`LightScratches`, `Clouding`, `Dirt`, `Dents`, `Whitening`, `Chipping`, `CornerWear`, `Creases`, `ShuffleBend`, `WaterDamage`, `Inking`, `Rips`, `BindersDents`
#### How the Grading Formula Works
Understanding this helps you know what training data matters most. The app grades cards as follows (from `ConditionEngine.swift`):
| Detected Damage | Grade Assigned |
| :--- | :--- |
| Any `Inking`, `Rips`, or `WaterDamage` detected | **Damaged** — immediately, regardless of anything else |
| 0 damage detections | **Near Mint (NM)** |
| 1–2 minor damage detections | **Excellent (EX)** |
| 3 or more minor damage detections | **Played (PL)** |
**Critical damage types** (`Inking`, `Rips`, `WaterDamage`) are the highest training priority — a single false positive will incorrectly grade a NM card as Damaged.
#### Materials List
| Item | Used For |
| :--- | :--- |
| 0000 (ultra-fine) steel wool | Surface scratches on foil cards |
| White vinyl eraser | Clouding/surface haze |
| Potting soil or cocoa powder | Dirt simulation |
| Ballpoint pen cap (rounded end) | Dents |
| Black Sharpie marker | Inking simulation |
| Spray bottle with water | Water damage |
| 3-ring binder | Binder dents |
| Rough mousepad or sandpaper | Corner wear |
| 50–100 bulk "draft chaff" cards | Cards to damage |
| Black felt or black paper | Background for edge photos |
| White card stock | Background for chipping photos |
| Desk lamp | Raking light source |
#### Sourcing Cards to Damage
Buy bulk worthless cards — do not damage your own collection.
- **eBay:** Search "MTG bulk commons lot" — 1000 cards for ~$10
- **TCGPlayer:** "Bulk commons" listings, often $0.01/card
- **Local game store:** Ask for "draft chaff" — often given away free
Also buy **pre-damaged cards** — natural damage looks more authentic to the model than simulated:
- **eBay:** Search "MTG damaged cards lot" or "heavily played bulk"
Aim for **50 cards per damage type** minimum. One card can be used for multiple damage types since each photo annotates only one damage area.
#### Raking Light Setup (Required for Surface and Structure Damage)
Most damage is invisible under flat overhead light. Raking light reveals it.
1. Place the card flat on a dark surface.
2. Position your desk lamp so light hits the card at a near-horizontal angle (5–15° above the surface) from one side.
3. The damage will cast visible shadows or catch the light clearly.
4. For scratches: slowly rotate the card until the scratches "light up" — photograph at that angle.
#### Damage Simulation Techniques
| Category | Damage Type | Folder Name | Simulation Technique | Photography Tip |
| :--- | :--- | :--- | :--- | :--- |
| **Surface** | Light Scratches | `LightScratches` | Rub foil surface gently with 0000 Steel Wool in one direction. | Raking light from the scratched direction. Rotate until scratches catch light. |
| **Surface** | Clouding | `Clouding` | Rub white vinyl eraser vigorously over foil surface in circles. | Diffused light. Compare side-by-side with a clean card for reference. |
| **Surface** | Dirt | `Dirt` | Press a damp fingertip into potting soil, then onto card surface. | Even lighting. Ensure dirt contrasts against the card art. |
| **Surface** | Dents | `Dents` | Press rounded end of a ballpoint pen cap firmly straight down. | Raking light at 10° to cast shadow inside the dent. |
| **Edges** | Whitening | `Whitening` | Rub card edges rapidly back and forth against denim jeans. | Black background. Macro close-up of the edge. |
| **Edges** | Chipping | `Chipping` | Use fingernail to carefully flake small pieces off the black border. | White background. Macro close-up. |
| **Edges** | Corner Wear | `CornerWear` | Rub corners against a rough mousepad with a circular motion. | Macro focus on the corner. Black background. |
| **Structure** | Creases | `Creases` | Fold corner sharply until a hard crease forms, then unfold. | Raking light to catch reflection off the crease ridge. |
| **Structure** | Shuffle Bend | `ShuffleBend` | Riffle shuffle the card aggressively 10+ times to create an arch. | Profile/side view to show curvature clearly. |
| **Structure** | Water Damage | `WaterDamage` | Mist card lightly with spray bottle, wait 60 seconds, air dry flat. | Raking light to show rippled surface texture. |
| **Critical** | Inking | `Inking` | Draw along whitened edges with black Sharpie to simulate edge touch-up. | UV/blacklight if available; otherwise strong white light at angle. |
| **Critical** | Rips | `Rips` | Tear edge slightly (~5mm). | High contrast background opposite to card border color. |
| **Critical** | Binder Dents | `BindersDents` | Press a 3-ring binder ring firmly into the card surface. | Raking light to show the circular crimp. |
#### What to Photograph per Damage Type
For each damage type, capture:
1. **30–50 cards showing that damage clearly** — positive training examples
2. **10–20 completely clean (undamaged) cards** — include these in every subfolder so the model learns the baseline
When annotating in Create ML, draw the bounding box **tightly around the damaged area only**. For shuffle bends, annotate the center of the arch. For edge damage, annotate the specific section of edge that is damaged, not the entire edge.
### Step 4: The "Edge Case" Validation List
Acquire these specific cheap cards to verify the logic-based detectors. **Note:** These are for **Manual Verification** (testing the app), not for Create ML training folders.