image-stitch（图像拼接）

Image Stitch Skill

Stitch multiple scrolling screenshots into one seamless image.

Step 1: Ask Stitch Direction

Use the Question tool to ask user ONE question only:

Question: Select stitch direction Options (exactly 2):

1. Vertical (top to bottom) - (Recommended) For vertically scrolling screenshots
2. Horizontal (left to right) - For horizontally scrolling screenshots

Step 2: Ask Overlap Size

Use the Question tool to ask user ONE question only:

Question: How much overlap between screenshots? Options (exactly 3):

1. Small overlap - Screenshots have minimal overlap (~10-20%)
2. Medium overlap - (Recommended) Screenshots have moderate overlap (~20-40%)
3. Large overlap - Screenshots have significant overlap (~40%+)

Based on user choice, set the edge parameter for modifying stitch.py:

• Small: EDGE_PARAM=150
• Medium: EDGE_PARAM=300
• Large: EDGE_PARAM=600

Step 3: Setup Environment

CRITICAL: Must run this BEFORE any stitching. The skill directory contains a venv that needs activation.

# Get skill directory (where stitch.py lives)
SKILL_DIR="/path/to/image-stitch"  # Replace with actual skill path

# Activate venv and install dependencies (idempotent)
cd "$SKILL_DIR" && \
python3 -m venv venv 2>/dev/null || true && \
source venv/bin/activate && \
pip install -q opencv-python numpy

# Modify stitch.py edge parameters based on user's overlap choice
EDGE_PARAM=300  # Set this based on Step 2 user choice
sd "edge_h = min\([0-9]+, h1 // 4, h2 // 4\)" "edge_h = min($EDGE_PARAM, h1 // 4, h2 // 4)" stitch.py
sd "edge_w = min\([0-9]+, w1 // 4, w2 // 4\)" "edge_w = min($EDGE_PARAM, w1 // 4, w2 // 4)" stitch.py

Step 4: Prepare Task Folders

Create task folders with timestamp:

TIMESTAMP=$(date +%Y%m%d_%H%M%S)
mkdir -p "$SKILL_DIR/input/$TIMESTAMP"
mkdir -p "$SKILL_DIR/output/$TIMESTAMP"

Step 5: Copy and Rename Images

Copy images from source to input/$TIMESTAMP/, renaming by sequence order.

IMPORTANT: Use find command to handle paths with spaces and special characters correctly.

SOURCE_PATH="/path/to/source"  # User provided path

# Find and copy images, sorted by filename, renamed to 01.png, 02.png, etc.
find "$SOURCE_PATH" -maxdepth 1 -type f \( -iname "*.png" -o -iname "*.jpg" -o -iname "*.jpeg" \) | \
sort | \
nl -nrz -w2 | \
while read num file; do
    cp "$file" "$SKILL_DIR/input/$TIMESTAMP/${num}.png"
done

Verify copied files:

ls -la "$SKILL_DIR/input/$TIMESTAMP/"

Step 6: Execute Stitching

IMPORTANT: Must run with venv activated.

cd "$SKILL_DIR" && source venv/bin/activate && \
python stitch.py \
    -i "input/$TIMESTAMP" \
    -o "output/$TIMESTAMP/stitched.png" \
    --debug \
    [--horizontal]  # Add this flag if user chose horizontal

Step 7: Report Result

After stitching, report to user with full absolute path:

Stitching complete!

Task ID: $TIMESTAMP
Input: $SKILL_DIR/input/$TIMESTAMP/ (<N> images)
Output: $SKILL_DIR/output/$TIMESTAMP/stitched.png (<W>x<H>)

Full output path:
$SKILL_DIR/output/$TIMESTAMP/stitched.png

Step 8: Ask to Open Output Folder

Use the Question tool to ask user ONE question only:

Question: Open output folder in Finder? Options (exactly 2):

1. Yes - Open the output folder
2. No - Skip opening folder

If user chooses "Yes", run:

open "$SKILL_DIR/output/$TIMESTAMP"

Script Options

Troubleshooting

ModuleNotFoundError: No module named 'cv2'

Cause: venv not activated or dependencies not installed. Fix: Run Step 3 (Setup Environment) again.

File copy fails with "No such file or directory"

Cause: Paths with spaces or special characters not quoted properly. Fix: Always use find with proper quoting as shown in Step 5.

No matches found: *.{png,jpg}

Cause: Brace expansion {a,b} not supported in all shells. Fix: Use find with -iname instead of glob patterns.

How It Works

1. Feature Detection: Uses ORB to find keypoints in overlap regions
2. Matching: Finds corresponding points between consecutive images
3. Offset Calculation:

- Primary axis: overlap amount (how much images share) - Secondary axis: alignment shift (corrects misalignment)

1. Stitching: Places images on canvas with calculated offsets
2. Cropping: Trims to common region

Requirements

• Python 3.10+
• opencv-python
• numpy

Dependencies are auto-installed in venv during Step 2.

Limitations

• Maximum 10 images per stitch
• Images should have 20%+ overlap for reliable matching
• Works best with content-rich areas (not pure solid colors)