MogensR's picture
Update processing/segmentation/hair_segmentation.py
bcf48d0
raw
history blame
21.2 kB
"""
Professional Hair Segmentation Module
=====================================
This module provides high-quality hair segmentation for video processing
using SAM2 + MatAnyone pipeline with comprehensive error handling and fallbacks.
Author: BackgroundFX Pro
License: MIT
"""
import os
import torch
import cv2
import numpy as np
import logging
from typing import Dict, List, Tuple, Optional, Union
from pathlib import Path
import warnings
from dataclasses import dataclass
from abc import ABC, abstractmethod
# Fix threading issues immediately
os.environ['OMP_NUM_THREADS'] = '4'
os.environ['MKL_NUM_THREADS'] = '4'
# Configure logging
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)
@dataclass
class SegmentationResult:
"""Result container for hair segmentation"""
mask: np.ndarray
confidence: float
coverage_percent: float
asymmetry_score: float
processing_time: float
fallback_used: bool
quality_score: float
error_message: Optional[str] = None
class BaseSegmentationModel(ABC):
"""Abstract base class for segmentation models"""
@abstractmethod
def initialize(self) -> bool:
"""Initialize the model"""
pass
@abstractmethod
def segment(self, frame: np.ndarray) -> np.ndarray:
"""Segment hair in frame"""
pass
@abstractmethod
def get_model_name(self) -> str:
"""Get model name for logging"""
pass
class SAM2Model(BaseSegmentationModel):
"""SAM2 segmentation model wrapper"""
def __init__(self, model_path: Optional[str] = None, device: str = 'auto'):
self.model_path = model_path
self.device = self._get_best_device(device)
self.predictor = None
self.initialized = False
def _get_best_device(self, device: str) -> str:
"""Determine best available device"""
if device == 'auto':
return 'cuda' if torch.cuda.is_available() else 'cpu'
return device
def initialize(self) -> bool:
"""Initialize SAM2 model"""
try:
logger.info("🤖 Initializing SAM2 model...")
# Import SAM2 (handle different installation methods)
try:
from sam2.build_sam import build_sam2_video_predictor
except ImportError:
logger.error("SAM2 not found. Please install SAM2.")
return False
# Build predictor
if self.model_path and Path(self.model_path).exists():
self.predictor = build_sam2_video_predictor(self.model_path, device=self.device)
else:
# Use default model
self.predictor = build_sam2_video_predictor("sam2_hiera_large.pt", device=self.device)
self.initialized = True
logger.info(f"✅ SAM2 initialized on {self.device}")
return True
except Exception as e:
logger.error(f"❌ SAM2 initialization failed: {e}")
return False
def segment(self, frame: np.ndarray) -> np.ndarray:
"""Segment using SAM2"""
if not self.initialized:
raise RuntimeError("SAM2 model not initialized")
try:
# Convert BGR to RGB
if len(frame.shape) == 3:
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
else:
frame_rgb = frame
# Set image for SAM2
self.predictor.set_image(frame_rgb)
# Auto-detect person in center (you can make this more sophisticated)
height, width = frame_rgb.shape[:2]
center_point = np.array([[width//2, height//2]])
# Predict mask
masks, scores, _ = self.predictor.predict(
point_coords=center_point,
point_labels=np.array([1])
)
# Return best mask
if len(masks) > 0:
best_mask_idx = np.argmax(scores)
return masks[best_mask_idx].astype(np.float32)
else:
return np.zeros((height, width), dtype=np.float32)
except Exception as e:
logger.error(f"SAM2 segmentation failed: {e}")
raise
def get_model_name(self) -> str:
return "SAM2"
class MatAnyoneModel(BaseSegmentationModel):
"""MatAnyone model wrapper with quality checking"""
def __init__(self, use_hf_api: bool = True, hf_token: Optional[str] = None):
self.use_hf_api = use_hf_api
self.hf_token = hf_token
self.client = None
self.processor = None
self.initialized = False
self.quality_threshold = 0.3
def initialize(self) -> bool:
"""Initialize MatAnyone model"""
try:
logger.info("🎭 Initializing MatAnyone model...")
if self.use_hf_api:
from gradio_client import Client
self.client = Client("PeiqingYang/MatAnyone", hf_token=self.hf_token)
logger.info("✅ MatAnyone HF API initialized")
else:
# Local MatAnyone initialization would go here
logger.warning("Local MatAnyone not implemented yet")
return False
self.initialized = True
return True
except Exception as e:
logger.error(f"❌ MatAnyone initialization failed: {e}")
return False
def segment(self, frame: np.ndarray) -> np.ndarray:
"""MatAnyone is primarily for matting, not segmentation"""
raise NotImplementedError("MatAnyone is used for matting, not direct segmentation")
def matte(self, image: np.ndarray, trimap: np.ndarray) -> np.ndarray:
"""Apply matting using MatAnyone"""
if not self.initialized:
raise RuntimeError("MatAnyone model not initialized")
try:
# Save temporary files
import tempfile
with tempfile.NamedTemporaryFile(suffix='.png', delete=False) as img_file:
cv2.imwrite(img_file.name, image)
img_path = img_file.name
with tempfile.NamedTemporaryFile(suffix='.png', delete=False) as tri_file:
cv2.imwrite(tri_file.name, trimap)
tri_path = tri_file.name
# Process with MatAnyone
if self.use_hf_api:
result = self._process_hf_api(img_path, tri_path)
else:
result = self._process_local(img_path, tri_path)
# Cleanup temp files
os.unlink(img_path)
os.unlink(tri_path)
return result
except Exception as e:
logger.error(f"MatAnyone matting failed: {e}")
raise
def _process_hf_api(self, image_path: str, trimap_path: str) -> np.ndarray:
"""Process using HuggingFace API"""
try:
result = self.client.predict(
image=image_path,
trimap=trimap_path,
api_name="/predict"
)
# Load result
if isinstance(result, str):
result_image = cv2.imread(result)
return result_image
else:
return result
except Exception as e:
logger.error(f"HF API processing failed: {e}")
raise
def _process_local(self, image_path: str, trimap_path: str) -> np.ndarray:
"""Process locally - placeholder for implementation"""
raise NotImplementedError("Local MatAnyone processing not implemented")
def get_model_name(self) -> str:
return "MatAnyone"
class TraditionalCVModel(BaseSegmentationModel):
"""Traditional computer vision fallback"""
def __init__(self):
self.initialized = False
def initialize(self) -> bool:
"""Initialize traditional CV methods"""
self.initialized = True
return True
def segment(self, frame: np.ndarray) -> np.ndarray:
"""Traditional hair segmentation using color and texture"""
try:
# Convert to different color spaces
hsv = cv2.cvtColor(frame, cv2.COLOR_BGR2HSV)
lab = cv2.cvtColor(frame, cv2.COLOR_BGR2LAB)
# Hair color detection
hair_mask_hsv = self._detect_hair_hsv(hsv)
hair_mask_lab = self._detect_hair_lab(lab)
# Combine masks
combined_mask = cv2.bitwise_or(hair_mask_hsv, hair_mask_lab)
# Morphological operations (using OpenCV instead of skimage)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (7, 7))
combined_mask = cv2.morphologyEx(combined_mask, cv2.MORPH_CLOSE, kernel)
combined_mask = cv2.morphologyEx(combined_mask, cv2.MORPH_OPEN, kernel)
return combined_mask.astype(np.float32) / 255.0
except Exception as e:
logger.error(f"Traditional CV segmentation failed: {e}")
raise
def _detect_hair_hsv(self, hsv: np.ndarray) -> np.ndarray:
"""Detect hair in HSV color space"""
# Multiple hair color ranges
ranges = [
# Dark hair
([0, 0, 0], [180, 255, 80]),
# Brown hair
([8, 50, 20], [25, 255, 200]),
# Blonde hair
([15, 30, 100], [35, 255, 255])
]
masks = []
for lower, upper in ranges:
mask = cv2.inRange(hsv, np.array(lower), np.array(upper))
masks.append(mask)
# Combine all color ranges
final_mask = masks[0]
for mask in masks[1:]:
final_mask = cv2.bitwise_or(final_mask, mask)
return final_mask
def _detect_hair_lab(self, lab: np.ndarray) -> np.ndarray:
"""Detect hair in LAB color space"""
l_channel = lab[:, :, 0]
hair_mask = cv2.inRange(l_channel, 0, 120)
return hair_mask
def get_model_name(self) -> str:
return "TraditionalCV"
class TemporalSmoother:
"""Temporal smoothing for video sequences"""
def __init__(self, smoothing_factor: float = 0.7, change_threshold: float = 0.05):
self.smoothing_factor = smoothing_factor
self.change_threshold = change_threshold
self.previous_mask = None
self.correction_count = 0
self.total_frames = 0
def smooth(self, current_mask: np.ndarray) -> Tuple[np.ndarray, bool]:
"""Apply temporal smoothing"""
self.total_frames += 1
corrected = False
if self.previous_mask is not None:
# Calculate change
diff = np.mean(np.abs(current_mask - self.previous_mask))
if diff > self.change_threshold:
# Apply smoothing
smoothed_mask = (self.smoothing_factor * current_mask +
(1 - self.smoothing_factor) * self.previous_mask)
self.correction_count += 1
corrected = True
else:
smoothed_mask = current_mask
else:
smoothed_mask = current_mask
self.previous_mask = smoothed_mask.copy()
return smoothed_mask, corrected
def get_correction_ratio(self) -> float:
"""Get ratio of frames that needed correction"""
return self.correction_count / max(self.total_frames, 1)
class HairSegmentationPipeline:
"""Main hair segmentation pipeline with multiple models and fallbacks"""
def __init__(self, config: Optional[Dict] = None):
self.config = config or {}
self.models = {}
self.active_model = None
self.fallback_models = []
self.temporal_smoother = TemporalSmoother()
self.initialized = False
# Setup models
self._setup_models()
def _setup_models(self):
"""Setup available models"""
try:
# Primary model: SAM2
sam2_model = SAM2Model(
model_path=self.config.get('sam2_model_path'),
device=self.config.get('device', 'auto')
)
self.models['sam2'] = sam2_model
# MatAnyone for matting
matanyone_model = MatAnyoneModel(
use_hf_api=self.config.get('use_hf_api', True),
hf_token=self.config.get('hf_token')
)
self.models['matanyone'] = matanyone_model
# Fallback: Traditional CV
traditional_model = TraditionalCVModel()
self.models['traditional'] = traditional_model
except Exception as e:
logger.error(f"Model setup failed: {e}")
def initialize(self, preferred_model: str = 'sam2') -> bool:
"""Initialize the pipeline"""
logger.info("🚀 Initializing Hair Segmentation Pipeline...")
# Try to initialize preferred model
if preferred_model in self.models:
if self.models[preferred_model].initialize():
self.active_model = preferred_model
logger.info(f"✅ Primary model {preferred_model} initialized")
else:
logger.warning(f"⚠️ Primary model {preferred_model} failed")
# Initialize fallback models
for model_name, model in self.models.items():
if model_name != self.active_model:
if model.initialize():
self.fallback_models.append(model_name)
logger.info(f"✅ Fallback model {model_name} ready")
# Check if we have at least one working model
if self.active_model or self.fallback_models:
self.initialized = True
logger.info(f"🎯 Pipeline ready - Active: {self.active_model}, Fallbacks: {self.fallback_models}")
return True
else:
logger.error("❌ No working models available")
return False
def segment_frame(self, frame: np.ndarray,
apply_temporal_smoothing: bool = True) -> SegmentationResult:
"""Segment hair in a single frame"""
if not self.initialized:
raise RuntimeError("Pipeline not initialized")
import time
start_time = time.time()
# Try active model first
mask, model_used, error_msg = self._try_segment_with_model(frame, self.active_model)
# If failed, try fallback models
if mask is None:
for fallback_model in self.fallback_models:
mask, model_used, error_msg = self._try_segment_with_model(frame, fallback_model)
if mask is not None:
break
if mask is None:
# Complete failure - return empty mask
h, w = frame.shape[:2]
mask = np.zeros((h, w), dtype=np.float32)
model_used = "none"
error_msg = "All models failed"
# Apply temporal smoothing
corrected = False
if apply_temporal_smoothing:
mask, corrected = self.temporal_smoother.smooth(mask)
# Calculate metrics
processing_time = time.time() - start_time
confidence = self._calculate_confidence(mask)
coverage = self._calculate_coverage(mask)
asymmetry = self._calculate_asymmetry(mask)
quality = self._calculate_quality(mask)
return SegmentationResult(
mask=mask,
confidence=confidence,
coverage_percent=coverage,
asymmetry_score=asymmetry,
processing_time=processing_time,
fallback_used=(model_used != self.active_model),
quality_score=quality,
error_message=error_msg
)
def _try_segment_with_model(self, frame: np.ndarray, model_name: str) -> Tuple[Optional[np.ndarray], str, Optional[str]]:
"""Try to segment with a specific model"""
if model_name not in self.models:
return None, model_name, f"Model {model_name} not available"
try:
mask = self.models[model_name].segment(frame)
return mask, model_name, None
except Exception as e:
error_msg = f"Model {model_name} failed: {str(e)}"
logger.warning(error_msg)
return None, model_name, error_msg
def _calculate_confidence(self, mask: np.ndarray) -> float:
"""Calculate mask confidence using OpenCV instead of skimage"""
# Edge sharpness
edges = cv2.Canny((mask * 255).astype(np.uint8), 50, 150)
edge_ratio = np.sum(edges > 0) / mask.size
# Mask smoothness using OpenCV Sobel instead of skimage gradient
grad_x = cv2.Sobel(mask, cv2.CV_64F, 1, 0, ksize=3)
grad_y = cv2.Sobel(mask, cv2.CV_64F, 0, 1, ksize=3)
gradient_magnitude = np.sqrt(grad_x**2 + grad_y**2)
smoothness = 1.0 / (1.0 + np.std(gradient_magnitude))
return min(edge_ratio * 0.3 + smoothness * 0.7, 1.0)
def _calculate_coverage(self, mask: np.ndarray) -> float:
"""Calculate hair coverage percentage"""
return (np.sum(mask > 0.5) / mask.size) * 100
def _calculate_asymmetry(self, mask: np.ndarray) -> float:
"""Calculate left-right asymmetry score"""
h, w = mask.shape[:2]
center_x = w // 2
left_half = mask[:, :center_x]
right_half = np.fliplr(mask[:, center_x:])
min_width = min(left_half.shape[1], right_half.shape[1])
left_half = left_half[:, :min_width]
right_half = right_half[:, :min_width]
return np.mean(np.abs(left_half - right_half))
def _calculate_quality(self, mask: np.ndarray) -> float:
"""Calculate overall mask quality"""
# Combine multiple quality metrics
confidence = self._calculate_confidence(mask)
coverage = self._calculate_coverage(mask) / 100.0
asymmetry_penalty = 1.0 - min(self._calculate_asymmetry(mask), 1.0)
return (confidence * 0.5 + coverage * 0.3 + asymmetry_penalty * 0.2)
def get_pipeline_stats(self) -> Dict:
"""Get pipeline performance statistics"""
return {
'active_model': self.active_model,
'fallback_models': self.fallback_models,
'temporal_correction_ratio': self.temporal_smoother.get_correction_ratio(),
'total_frames_processed': self.temporal_smoother.total_frames,
'corrections_applied': self.temporal_smoother.correction_count
}
# Convenience functions
def create_pipeline(config: Optional[Dict] = None) -> HairSegmentationPipeline:
"""Create and initialize hair segmentation pipeline"""
pipeline = HairSegmentationPipeline(config)
pipeline.initialize()
return pipeline
def segment_image(image_path: str, config: Optional[Dict] = None) -> SegmentationResult:
"""Segment hair in a single image"""
pipeline = create_pipeline(config)
frame = cv2.imread(image_path)
return pipeline.segment_frame(frame)
def segment_video_frames(video_frames: List[np.ndarray],
config: Optional[Dict] = None) -> List[SegmentationResult]:
"""Segment hair in multiple video frames"""
pipeline = create_pipeline(config)
results = []
for frame in video_frames:
result = pipeline.segment_frame(frame)
results.append(result)
return results
# Example usage
if __name__ == "__main__":
# Example configuration
config = {
'sam2_model_path': None, # Use default
'device': 'auto',
'use_hf_api': True,
'hf_token': None # Set your token if needed
}
# Create pipeline
pipeline = create_pipeline(config)
# Test with example frame (you would load your actual frame)
test_frame = np.random.randint(0, 255, (480, 640, 3), dtype=np.uint8)
# Segment frame
result = pipeline.segment_frame(test_frame)
# Print results
print(f"Segmentation Results:")
print(f" Coverage: {result.coverage_percent:.1f}%")
print(f" Confidence: {result.confidence:.3f}")
print(f" Quality: {result.quality_score:.3f}")
print(f" Processing time: {result.processing_time:.2f}s")
print(f" Fallback used: {result.fallback_used}")
# Get pipeline stats
stats = pipeline.get_pipeline_stats()
print(f"\nPipeline Stats:")
print(f" Active model: {stats['active_model']}")
print(f" Fallbacks: {stats['fallback_models']}")
print(f" Correction ratio: {stats['temporal_correction_ratio']:.3f}")