ring-sizer / src /finger_segmentation.py
feng-x's picture
Upload folder using huggingface_hub
347d1a8 verified
"""
Hand and finger segmentation utilities.
This module handles:
- Hand detection using MediaPipe
- Hand mask generation
- Individual finger isolation
- Mask cleanup and validation
"""
import cv2
import numpy as np
from typing import Optional, Dict, Any, Literal, List, Tuple
import mediapipe as mp
from mediapipe.tasks import python
from mediapipe.tasks.python import vision
import urllib.request
import os
from pathlib import Path
# Import debug observer and drawing functions
from src.debug_observer import (
DebugObserver,
draw_landmarks_overlay,
draw_hand_skeleton,
draw_detection_info,
)
FingerIndex = Literal["auto", "index", "middle", "ring", "pinky"]
# MediaPipe hand landmark indices for each finger
# Each finger has 4 landmarks: MCP (knuckle), PIP, DIP, TIP
FINGER_LANDMARKS = {
"index": [5, 6, 7, 8],
"middle": [9, 10, 11, 12],
"ring": [13, 14, 15, 16],
"pinky": [17, 18, 19, 20],
}
# Thumb landmarks (special case - not typically used for ring measurement)
THUMB_LANDMARKS = [1, 2, 3, 4]
# Wrist landmark
WRIST_LANDMARK = 0
# Palm landmarks (for creating hand mask)
PALM_LANDMARKS = [0, 1, 5, 9, 13, 17]
# Model path
MODEL_PATH = os.path.join(os.path.dirname(__file__), "..", ".model", "hand_landmarker.task")
MODEL_URL = "https://storage.googleapis.com/mediapipe-models/hand_landmarker/hand_landmarker/float16/1/hand_landmarker.task"
# Initialize MediaPipe Hands (lazy loading)
_hands_detector = None
def _download_model():
"""Download the hand landmarker model if not present."""
if not os.path.exists(MODEL_PATH):
os.makedirs(os.path.dirname(MODEL_PATH), exist_ok=True)
print(f"Downloading hand landmarker model...")
urllib.request.urlretrieve(MODEL_URL, MODEL_PATH)
print(f"Model downloaded to {MODEL_PATH}")
def _get_hands_detector(force_new: bool = False):
"""Get or initialize the MediaPipe Hands detector."""
global _hands_detector
if _hands_detector is None or force_new:
_download_model()
base_options = python.BaseOptions(model_asset_path=MODEL_PATH)
options = vision.HandLandmarkerOptions(
base_options=base_options,
num_hands=2,
min_hand_detection_confidence=0.3, # Lower threshold for better detection
min_tracking_confidence=0.3,
)
_hands_detector = vision.HandLandmarker.create_from_options(options)
return _hands_detector
def _try_detect_hand(detector, image: np.ndarray) -> Optional[Tuple[Any, int]]:
"""
Try to detect hand in image, returns (results, rotation_code) or None.
rotation_code: 0=none, 1=90cw, 2=180, 3=90ccw
"""
# Try different rotations to handle various image orientations
rotations = [
(image, 0),
(cv2.rotate(image, cv2.ROTATE_90_CLOCKWISE), 1),
(cv2.rotate(image, cv2.ROTATE_90_COUNTERCLOCKWISE), 3),
(cv2.rotate(image, cv2.ROTATE_180), 2),
]
best_result = None
best_confidence = 0
best_rotation = 0
for rotated, rot_code in rotations:
# Convert to RGB and ensure contiguous memory layout
rgb = cv2.cvtColor(rotated, cv2.COLOR_BGR2RGB)
rgb = np.ascontiguousarray(rgb)
mp_image = mp.Image(image_format=mp.ImageFormat.SRGB, data=rgb)
results = detector.detect(mp_image)
if results.hand_landmarks:
# Get best confidence among detected hands
for i, handedness in enumerate(results.handedness):
conf = handedness[0].score
if conf > best_confidence:
best_confidence = conf
best_result = results
best_rotation = rot_code
if best_result is None:
return None
return best_result, best_rotation
def _transform_landmarks_for_rotation(
landmarks: np.ndarray,
rotation_code: int,
original_h: int,
original_w: int,
) -> np.ndarray:
"""
Transform landmarks from rotated coordinates back to original image coordinates.
"""
if rotation_code == 0:
# No rotation
return landmarks
elif rotation_code == 1:
# Was rotated 90 CW, so transform back (90 CCW)
# In rotated: (x, y) with size (h, w) -> original: (y, w-1-x) with size (w, h)
new_landmarks = np.zeros_like(landmarks)
new_landmarks[:, 0] = landmarks[:, 1] * original_w # y -> x
new_landmarks[:, 1] = (1 - landmarks[:, 0]) * original_h # (1-x) -> y
return new_landmarks
elif rotation_code == 2:
# Was rotated 180
new_landmarks = np.zeros_like(landmarks)
new_landmarks[:, 0] = (1 - landmarks[:, 0]) * original_w
new_landmarks[:, 1] = (1 - landmarks[:, 1]) * original_h
return new_landmarks
elif rotation_code == 3:
# Was rotated 90 CCW, so transform back (90 CW)
new_landmarks = np.zeros_like(landmarks)
new_landmarks[:, 0] = (1 - landmarks[:, 1]) * original_w
new_landmarks[:, 1] = landmarks[:, 0] * original_h
return new_landmarks
return landmarks
def detect_hand_orientation(
landmarks_normalized: np.ndarray,
finger: FingerIndex = "index"
) -> float:
"""
Detect hand orientation angle from vertical (canonical orientation).
Canonical orientation: wrist at bottom, fingers pointing upward.
Args:
landmarks_normalized: MediaPipe hand landmarks (21x2) in normalized [0-1] coordinates
finger: Which finger to use for orientation detection (default: "index")
Returns:
Angle in degrees to rotate image clockwise to achieve canonical orientation.
Returns one of: 0, 90, 180, 270
"""
# Get wrist (landmark 0) and specified finger tip
wrist = landmarks_normalized[WRIST_LANDMARK]
# Use specified finger, fallback to middle if invalid
if finger in FINGER_LANDMARKS:
finger_tip = landmarks_normalized[FINGER_LANDMARKS[finger][3]]
else:
# Fallback to middle finger for "auto" or invalid values
finger_tip = landmarks_normalized[FINGER_LANDMARKS["middle"][3]]
# Compute vector from wrist to fingertip
direction = finger_tip - wrist
# Compute angle from vertical upward direction
# In image coordinates: y increases downward, x increases rightward
# Vertical upward = (0, -1) in (x, y)
# angle = atan2(cross, dot) where cross = dx*(-1) - dy*0, dot = dx*0 + dy*(-1)
angle_rad = np.arctan2(direction[0], -direction[1])
angle_deg = angle_rad * 180.0 / np.pi
# angle_deg is now in range [-180, 180]:
# 0° = fingers pointing up (canonical)
# 90° = fingers pointing right
# 180° = fingers pointing down
# -90° = fingers pointing left
# Convert to [0, 360] range
if angle_deg < 0:
angle_deg += 360
# Snap to nearest 90° increment
# We want to return how much to rotate CW to get to canonical (0°)
rotation_needed = _snap_to_orthogonal(angle_deg)
return rotation_needed
def _snap_to_orthogonal(angle_deg: float) -> int:
"""
Snap angle to nearest orthogonal rotation (0, 90, 180, 270).
Args:
angle_deg: Angle in degrees [0, 360]
Returns:
Rotation needed in degrees (0, 90, 180, 270) to rotate CW to canonical orientation
"""
# If angle is 0±45°, no rotation needed
# If angle is 90±45°, need to rotate 270° CW (or 90° CCW) to get to 0°
# If angle is 180±45°, need to rotate 180°
# If angle is 270±45°, need to rotate 90° CW
# Determine which quadrant (with 45° tolerance)
if angle_deg < 45 or angle_deg >= 315:
return 0 # Already upright
elif 45 <= angle_deg < 135:
return 270 # Pointing right, rotate 270° CW (= 90° CCW)
elif 135 <= angle_deg < 225:
return 180 # Upside down, rotate 180°
else: # 225 <= angle_deg < 315
return 90 # Pointing left, rotate 90° CW
def normalize_hand_orientation(
image: np.ndarray,
landmarks_normalized: np.ndarray,
finger: FingerIndex = "index",
) -> Tuple[np.ndarray, int]:
"""
Rotate image to canonical hand orientation (wrist at bottom, fingers up).
Args:
image: Input BGR image
landmarks_normalized: MediaPipe landmarks in normalized [0-1] coordinates
finger: Which finger to use for orientation detection (default: "index")
Returns:
Tuple of (rotated_image, rotation_angle_degrees)
rotation_angle_degrees is one of: 0, 90, 180, 270
"""
# Detect hand orientation based on specified finger
rotation_needed = detect_hand_orientation(landmarks_normalized, finger)
# Rotate image if needed
if rotation_needed == 0:
return image, 0
elif rotation_needed == 90:
return cv2.rotate(image, cv2.ROTATE_90_CLOCKWISE), 90
elif rotation_needed == 180:
return cv2.rotate(image, cv2.ROTATE_180), 180
elif rotation_needed == 270:
return cv2.rotate(image, cv2.ROTATE_90_COUNTERCLOCKWISE), 270
else:
# Shouldn't happen, but return original as fallback
print(f"Warning: Unexpected rotation angle {rotation_needed}, skipping rotation")
return image, 0
def segment_hand(
image: np.ndarray,
finger: FingerIndex = "index",
max_dimension: int = 1280,
debug_dir: Optional[str] = None,
) -> Optional[Dict[str, Any]]:
"""
Detect and segment hand from image using MediaPipe.
Args:
image: Input BGR image
finger: Which finger to use for orientation detection (default: "index")
max_dimension: Maximum dimension for processing (large images are resized)
debug_dir: Optional directory to save debug images
Returns:
Dictionary containing:
- landmarks: 21x2 array of landmark positions (pixel coordinates)
- landmarks_normalized: 21x2 array of normalized coordinates [0-1]
- mask: Binary hand mask
- confidence: Detection confidence
- handedness: "Left" or "Right"
Or None if no hand detected
"""
# Create debug observer if debug mode enabled
observer = DebugObserver(debug_dir) if debug_dir else None
h, w = image.shape[:2]
# Debug: Save original image
if observer:
observer.save_stage("01_original", image)
# Resize if image is too large (MediaPipe works better with smaller images)
scale = 1.0
if max(h, w) > max_dimension:
scale = max_dimension / max(h, w)
new_w = int(w * scale)
new_h = int(h * scale)
resized = cv2.resize(image, (new_w, new_h), interpolation=cv2.INTER_AREA)
else:
resized = image
new_h, new_w = h, w
# Debug: Save resized image (if resized)
if scale != 1.0 and observer:
observer.save_stage("02_resized_for_detection", resized)
# Process with MediaPipe (try multiple rotations)
detector = _get_hands_detector()
detection_result = _try_detect_hand(detector, resized)
if detection_result is None:
return None
results, rotation_code = detection_result
# Select the best hand (highest confidence)
best_hand_idx = 0
best_conf = 0
for i, handedness in enumerate(results.handedness):
if handedness[0].score > best_conf:
best_conf = handedness[0].score
best_hand_idx = i
hand_landmarks = results.hand_landmarks[best_hand_idx]
handedness = results.handedness[best_hand_idx]
# Extract landmark coordinates (normalized 0-1 in rotated image)
landmarks_normalized_rotated = np.array([
[lm.x, lm.y] for lm in hand_landmarks
])
# NEW: Normalize hand orientation to canonical (wrist at bottom, fingers up)
# This is done in the detected-rotation space first
if rotation_code == 1:
# Was rotated 90 CW
rotated_image = cv2.rotate(resized, cv2.ROTATE_90_CLOCKWISE)
elif rotation_code == 2:
# Was rotated 180
rotated_image = cv2.rotate(resized, cv2.ROTATE_180)
elif rotation_code == 3:
# Was rotated 90 CCW
rotated_image = cv2.rotate(resized, cv2.ROTATE_90_COUNTERCLOCKWISE)
else:
rotated_image = resized
# Now normalize orientation based on hand direction
canonical_image, orientation_rotation = normalize_hand_orientation(
rotated_image, landmarks_normalized_rotated, finger
)
# Update landmarks for orientation normalization
if orientation_rotation != 0:
rot_h, rot_w = rotated_image.shape[:2]
landmarks_px_rotated = landmarks_normalized_rotated.copy()
landmarks_px_rotated[:, 0] *= rot_w
landmarks_px_rotated[:, 1] *= rot_h
# Apply rotation transform to landmarks
if orientation_rotation == 90:
# Rotate 90 CW: (x, y) -> (h-1-y, x)
new_landmarks = np.zeros_like(landmarks_px_rotated)
new_landmarks[:, 0] = rot_h - 1 - landmarks_px_rotated[:, 1]
new_landmarks[:, 1] = landmarks_px_rotated[:, 0]
landmarks_px_canonical = new_landmarks
elif orientation_rotation == 180:
# Rotate 180: (x, y) -> (w-1-x, h-1-y)
new_landmarks = np.zeros_like(landmarks_px_rotated)
new_landmarks[:, 0] = rot_w - 1 - landmarks_px_rotated[:, 0]
new_landmarks[:, 1] = rot_h - 1 - landmarks_px_rotated[:, 1]
landmarks_px_canonical = new_landmarks
elif orientation_rotation == 270:
# Rotate 90 CCW: (x, y) -> (y, w-1-x)
new_landmarks = np.zeros_like(landmarks_px_rotated)
new_landmarks[:, 0] = landmarks_px_rotated[:, 1]
new_landmarks[:, 1] = rot_w - 1 - landmarks_px_rotated[:, 0]
landmarks_px_canonical = new_landmarks
else:
landmarks_px_canonical = landmarks_px_rotated
# Update normalized landmarks for canonical image
can_h, can_w = canonical_image.shape[:2]
landmarks_normalized_canonical = landmarks_px_canonical.copy()
landmarks_normalized_canonical[:, 0] /= can_w
landmarks_normalized_canonical[:, 1] /= can_h
else:
landmarks_normalized_canonical = landmarks_normalized_rotated
# Scale landmarks back to original resolution if needed
if scale != 1.0:
canonical_full = cv2.resize(canonical_image, (int(canonical_image.shape[1] / scale),
int(canonical_image.shape[0] / scale)),
interpolation=cv2.INTER_CUBIC)
else:
canonical_full = canonical_image
# Final landmarks in canonical full resolution
can_full_h, can_full_w = canonical_full.shape[:2]
landmarks_canonical = landmarks_normalized_canonical.copy()
landmarks_canonical[:, 0] *= can_full_w
landmarks_canonical[:, 1] *= can_full_h
# Debug: Draw landmarks overlay in canonical orientation
if observer:
observer.draw_and_save("03_landmarks_overlay_canonical", canonical_full,
draw_landmarks_overlay, landmarks_canonical, label=True)
observer.draw_and_save("04_hand_skeleton_canonical", canonical_full,
draw_hand_skeleton, landmarks_canonical)
observer.draw_and_save("05_detection_info_canonical", canonical_full,
draw_detection_info, handedness[0].score,
handedness[0].category_name,
f"det={rotation_code}, orient={orientation_rotation}")
# Generate hand mask at canonical resolution
mask = _create_hand_mask(landmarks_canonical, (can_full_h, can_full_w))
return {
"landmarks": landmarks_canonical,
"landmarks_normalized": landmarks_normalized_canonical,
"mask": mask,
"confidence": handedness[0].score,
"handedness": handedness[0].category_name,
"rotation_applied": rotation_code,
"orientation_rotation": orientation_rotation,
"canonical_image": canonical_full, # Return the canonical image for downstream processing
}
def _create_hand_mask(landmarks: np.ndarray, shape: Tuple[int, int]) -> np.ndarray:
"""
Create a binary mask of the hand region from landmarks.
Args:
landmarks: 21x2 array of landmark pixel coordinates
shape: (height, width) of output mask
Returns:
Binary mask (uint8, 0 or 255)
"""
h, w = shape
mask = np.zeros((h, w), dtype=np.uint8)
# Create convex hull of all landmarks
hull_points = cv2.convexHull(landmarks.astype(np.int32))
cv2.fillConvexPoly(mask, hull_points, 255)
# Also fill individual finger regions for better coverage
for finger_name, indices in FINGER_LANDMARKS.items():
finger_pts = landmarks[indices].astype(np.int32)
cv2.fillConvexPoly(mask, finger_pts, 255)
# Fill thumb
thumb_pts = landmarks[THUMB_LANDMARKS].astype(np.int32)
cv2.fillConvexPoly(mask, thumb_pts, 255)
# Apply morphological operations to smooth the mask
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (15, 15))
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel)
mask = cv2.morphologyEx(mask, cv2.MORPH_OPEN, kernel)
return mask
def _calculate_finger_extension(landmarks: np.ndarray, finger_indices: List[int]) -> float:
"""
Calculate how extended a finger is based on landmark positions.
Returns a score where higher = more extended.
"""
if len(finger_indices) < 4:
return 0.0
# Get finger landmarks
mcp = landmarks[finger_indices[0]] # Knuckle
pip = landmarks[finger_indices[1]] # First joint
dip = landmarks[finger_indices[2]] # Second joint
tip = landmarks[finger_indices[3]] # Fingertip
# Calculate vectors
mcp_to_tip = tip - mcp
mcp_to_pip = pip - mcp
# Extension score based on:
# 1. Distance from knuckle to tip (longer = more extended)
finger_length = np.linalg.norm(mcp_to_tip)
# 2. Straightness (how aligned are the joints)
pip_to_dip = dip - pip
dip_to_tip = tip - dip
# Dot products to check alignment (1 = straight, -1 = bent back)
if np.linalg.norm(mcp_to_pip) > 0 and np.linalg.norm(pip_to_dip) > 0:
align1 = np.dot(mcp_to_pip, pip_to_dip) / (np.linalg.norm(mcp_to_pip) * np.linalg.norm(pip_to_dip))
else:
align1 = 0
if np.linalg.norm(pip_to_dip) > 0 and np.linalg.norm(dip_to_tip) > 0:
align2 = np.dot(pip_to_dip, dip_to_tip) / (np.linalg.norm(pip_to_dip) * np.linalg.norm(dip_to_tip))
else:
align2 = 0
straightness = (align1 + align2) / 2
# Combined score
return finger_length * (0.5 + 0.5 * max(0, straightness))
def _create_finger_roi_mask(
finger_landmarks: np.ndarray,
all_landmarks: np.ndarray,
shape: Tuple[int, int],
expansion_factor: float = 1.8,
) -> np.ndarray:
"""
Create a Region of Interest (ROI) mask around finger landmarks.
This creates a generous bounding region that should contain the entire finger
without cutting off edges, but excludes other fingers.
Args:
finger_landmarks: 4x2 array of finger landmark positions (MCP, PIP, DIP, TIP)
all_landmarks: 21x2 array of all hand landmarks
shape: (height, width) of output mask
expansion_factor: How much to expand perpendicular to finger axis
Returns:
Binary ROI mask
"""
h, w = shape
roi_mask = np.zeros((h, w), dtype=np.uint8)
# Calculate finger axis direction
mcp = finger_landmarks[0]
tip = finger_landmarks[3]
finger_axis = tip - mcp
finger_length = np.linalg.norm(finger_axis)
if finger_length < 1:
return roi_mask
finger_direction = finger_axis / finger_length
# Perpendicular direction
perp = np.array([-finger_direction[1], finger_direction[0]])
# Estimate finger width from landmark spacing
# Use median distance between consecutive landmarks as width proxy
segment_lengths = []
for i in range(len(finger_landmarks) - 1):
seg_len = np.linalg.norm(finger_landmarks[i + 1] - finger_landmarks[i])
segment_lengths.append(seg_len)
avg_segment = np.median(segment_lengths) if segment_lengths else finger_length / 3
# Finger width is roughly 1/3 to 1/2 of segment length
base_width = avg_segment * 0.6 * expansion_factor
# Extend ROI slightly beyond landmarks (towards palm and beyond tip)
wrist = all_landmarks[WRIST_LANDMARK]
palm_direction = mcp - wrist
palm_direction = palm_direction / (np.linalg.norm(palm_direction) + 1e-8)
# Extend 20% beyond MCP toward palm
extended_base = mcp - palm_direction * finger_length * 0.2
# Extend 10% beyond tip
extended_tip = tip + finger_direction * finger_length * 0.1
# Create polygon along finger with wider margins
polygon_points = []
num_samples = 8 # More points for smoother ROI
for i in range(num_samples):
t = i / (num_samples - 1)
# Interpolate from extended base to extended tip
pt = extended_base + (extended_tip - extended_base) * t
# Width varies: wider at base, narrower at tip
width_scale = 1.0 - 0.2 * t
half_width = base_width * width_scale / 2
# Add left and right points
left = pt + perp * half_width
right = pt - perp * half_width
polygon_points.append((left, right))
# Build polygon
polygon = []
for left, right in polygon_points:
polygon.append(left)
for left, right in reversed(polygon_points):
polygon.append(right)
polygon = np.array(polygon, dtype=np.int32)
cv2.fillPoly(roi_mask, [polygon], 255)
return roi_mask
def _isolate_finger_from_hand_mask(
hand_mask: np.ndarray,
finger_landmarks: np.ndarray,
all_landmarks: np.ndarray,
min_area: int = 500,
) -> Optional[np.ndarray]:
"""
Isolate finger using pixel-level intersection of hand mask with finger ROI.
This is the preferred method as it preserves actual finger edges from MediaPipe
rather than creating a synthetic polygon.
Args:
hand_mask: Full hand mask from MediaPipe (pixel-accurate)
finger_landmarks: 4x2 array of finger landmarks
all_landmarks: 21x2 array of all hand landmarks
min_area: Minimum valid finger area
Returns:
Binary finger mask, or None if isolation fails
"""
h, w = hand_mask.shape
# Create ROI mask around finger
roi_mask = _create_finger_roi_mask(finger_landmarks, all_landmarks, (h, w))
# Intersect hand mask with finger ROI
# This preserves real pixel-level edges from MediaPipe
finger_mask = cv2.bitwise_and(hand_mask, roi_mask)
# Find connected components to remove fragments from other fingers
num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(
finger_mask, connectivity=8
)
if num_labels <= 1:
return None
# Select component closest to finger landmarks centroid
landmarks_centroid = np.mean(finger_landmarks, axis=0)
best_component = None
best_distance = float('inf')
for i in range(1, num_labels): # Skip background (0)
area = stats[i, cv2.CC_STAT_AREA]
if area < min_area:
continue
component_centroid = centroids[i]
dist = np.linalg.norm(component_centroid - landmarks_centroid)
if dist < best_distance:
best_distance = dist
best_component = i
if best_component is None:
return None
# Create final mask with only the selected component
final_mask = np.zeros_like(finger_mask)
final_mask[labels == best_component] = 255
return final_mask
def isolate_finger(
hand_data: Dict[str, Any],
finger: FingerIndex = "auto",
image_shape: Optional[Tuple[int, int]] = None,
) -> Optional[Dict[str, Any]]:
"""
Isolate a specific finger from hand segmentation data.
Args:
hand_data: Output from segment_hand()
finger: Which finger to isolate, or "auto" to select most extended
image_shape: (height, width) for mask generation
Returns:
Dictionary containing:
- mask: Binary finger mask
- landmarks: Finger landmark positions (4x2 array)
- base_point: Palm-side base of finger (MCP joint)
- tip_point: Fingertip position
- finger_name: Name of the isolated finger
Or None if finger cannot be isolated
"""
landmarks = hand_data["landmarks"]
if image_shape is None:
if "mask" in hand_data:
image_shape = hand_data["mask"].shape[:2]
else:
return None
# Determine which finger to use
if finger == "auto":
best_finger = None
best_score = -1
for finger_name, indices in FINGER_LANDMARKS.items():
score = _calculate_finger_extension(landmarks, indices)
if score > best_score:
best_score = score
best_finger = finger_name
if best_finger is None:
return None
finger = best_finger
if finger not in FINGER_LANDMARKS:
return None
indices = FINGER_LANDMARKS[finger]
finger_landmarks = landmarks[indices]
# Create finger mask using pixel-level approach (preferred)
mask = None
method_used = "unknown"
if "mask" in hand_data and hand_data["mask"] is not None:
mask = _isolate_finger_from_hand_mask(
hand_data["mask"],
finger_landmarks,
landmarks,
min_area=500,
)
if mask is not None:
method_used = "pixel-level"
print(f" Finger isolated using pixel-level segmentation")
else:
print(f" Pixel-level segmentation failed, falling back to polygon")
# Fallback to polygon-based approach
if mask is None:
mask = _create_finger_mask(landmarks, indices, image_shape)
if mask is not None:
method_used = "polygon"
print(f" Finger isolated using polygon-based segmentation (fallback)")
else:
print(f" Both segmentation methods failed")
return None
return {
"mask": mask,
"landmarks": finger_landmarks,
"base_point": finger_landmarks[0], # MCP joint
"tip_point": finger_landmarks[3], # Fingertip
"finger_name": finger,
"method": method_used,
}
def _create_finger_mask(
all_landmarks: np.ndarray,
finger_indices: List[int],
shape: Tuple[int, int],
width_factor: float = 2.5,
) -> Optional[np.ndarray]:
"""
Create a binary mask for a single finger using polygon approximation.
This is the fallback method when pixel-level segmentation fails.
Args:
all_landmarks: All 21 hand landmarks
finger_indices: Indices of the 4 finger landmarks
shape: (height, width) of output mask
width_factor: Multiplier for estimated finger width
Returns:
Binary mask of finger region
"""
h, w = shape
mask = np.zeros((h, w), dtype=np.uint8)
finger_landmarks = all_landmarks[finger_indices]
# Estimate finger width based on joint spacing
mcp_idx = finger_indices[0]
adjacent_distances = []
for other_finger, other_indices in FINGER_LANDMARKS.items():
other_mcp = other_indices[0]
if other_mcp != mcp_idx:
dist = np.linalg.norm(all_landmarks[mcp_idx] - all_landmarks[other_mcp])
adjacent_distances.append(dist)
if adjacent_distances:
estimated_width = min(adjacent_distances) * 0.4 * width_factor
else:
finger_length = np.linalg.norm(finger_landmarks[3] - finger_landmarks[0])
estimated_width = finger_length / 6 * width_factor
# Create polygon along finger with estimated width
polygon_points = []
for i in range(len(finger_landmarks)):
pt = finger_landmarks[i]
if i < len(finger_landmarks) - 1:
direction = finger_landmarks[i + 1] - pt
else:
direction = pt - finger_landmarks[i - 1]
perp = np.array([-direction[1], direction[0]])
perp_norm = np.linalg.norm(perp)
if perp_norm > 0:
perp = perp / perp_norm
width_scale = 1.0 - 0.3 * (i / (len(finger_landmarks) - 1))
half_width = estimated_width * width_scale / 2
left = pt + perp * half_width
right = pt - perp * half_width
polygon_points.append((left, right))
# Build polygon: go up left side, then down right side
polygon = []
for left, right in polygon_points:
polygon.append(left)
for left, right in reversed(polygon_points):
polygon.append(right)
polygon = np.array(polygon, dtype=np.int32)
cv2.fillPoly(mask, [polygon], 255)
# Extend mask slightly towards palm
mcp = finger_landmarks[0]
wrist = all_landmarks[WRIST_LANDMARK]
palm_direction = mcp - wrist
palm_direction = palm_direction / (np.linalg.norm(palm_direction) + 1e-8)
finger_length = np.linalg.norm(finger_landmarks[3] - finger_landmarks[0])
extension = palm_direction * finger_length * 0.15
extended_base = mcp - extension
perp = np.array([-palm_direction[1], palm_direction[0]])
half_width = estimated_width / 2
ext_polygon = np.array([
mcp + perp * half_width,
mcp - perp * half_width,
extended_base - perp * half_width * 0.8,
extended_base + perp * half_width * 0.8,
], dtype=np.int32)
cv2.fillPoly(mask, [ext_polygon], 255)
return mask
def clean_mask(
mask: np.ndarray,
min_area: int = 1000,
) -> Optional[np.ndarray]:
"""
Clean a binary mask by extracting largest component and applying morphology.
Args:
mask: Input binary mask
min_area: Minimum valid area in pixels
Returns:
Cleaned binary mask, or None if no valid component found
"""
if mask is None or mask.size == 0:
return None
# Find connected components
num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(mask, connectivity=8)
if num_labels <= 1:
return None
# Find largest component (excluding background at index 0)
largest_idx = 1
largest_area = 0
for i in range(1, num_labels):
area = stats[i, cv2.CC_STAT_AREA]
if area > largest_area:
largest_area = area
largest_idx = i
if largest_area < min_area:
return None
# Create mask with only the largest component
cleaned = np.zeros_like(mask)
cleaned[labels == largest_idx] = 255
# Apply morphological smoothing
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (7, 7))
cleaned = cv2.morphologyEx(cleaned, cv2.MORPH_CLOSE, kernel)
cleaned = cv2.morphologyEx(cleaned, cv2.MORPH_OPEN, kernel)
# Smooth edges with Gaussian blur and re-threshold
cleaned = cv2.GaussianBlur(cleaned, (5, 5), 0)
_, cleaned = cv2.threshold(cleaned, 127, 255, cv2.THRESH_BINARY)
return cleaned
def get_finger_contour(
mask: np.ndarray,
smooth: bool = True,
) -> Optional[np.ndarray]:
"""
Extract outer contour from finger mask.
Args:
mask: Binary finger mask
smooth: Whether to apply contour smoothing
Returns:
Contour points as Nx2 array, or None if no contour found
"""
if mask is None:
return None
# Find contours
contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if not contours:
return None
# Get the largest contour
largest_contour = max(contours, key=cv2.contourArea)
# Reshape to Nx2
contour = largest_contour.reshape(-1, 2)
if smooth and len(contour) > 10:
# Apply contour smoothing using approximation
epsilon = 0.005 * cv2.arcLength(largest_contour, True)
smoothed = cv2.approxPolyDP(largest_contour, epsilon, True)
contour = smoothed.reshape(-1, 2)
return contour.astype(np.float32)