feat: add debug mode, distance-based thresholds, and improve pose detection
- Add debug mode support for encoding statistics in API responses - Debug info includes encoding length, min/max/mean/std, and first 10 values - Frontend logs encoding stats to browser console when debug enabled - Identify page enables debug mode by default - Implement distance-based confidence thresholds for stricter matching - Borderline distances require higher confidence (70-95% vs 50%) - Applied when use_distance_based_thresholds=True (auto-match) - Reduces false positives for borderline matches - Dual tolerance system for auto-match - Default tolerance 0.6 for regular browsing (more lenient) - Run auto-match button uses 0.5 tolerance with distance-based thresholds (stricter) - Auto-accept threshold updated to 85% (from 70%) - Enhance pose detection with single-eye detection - Profile threshold reduced from 30° to 15° (stricter) - Detect single-eye visibility for extreme profile views - Infer profile direction from landmark visibility - Improved face width threshold (20px vs 10px) - Clean up debug code - Remove test photo UUID checks from production code - Remove debug print statements - Replace print statements with proper logging
This commit is contained in:
parent
6b6b1449b2
commit
a6ba78cd54
67
admin-frontend/public/enable-dev-mode.html
Normal file
67
admin-frontend/public/enable-dev-mode.html
Normal file
@ -0,0 +1,67 @@
|
||||
<!DOCTYPE html>
|
||||
<html>
|
||||
<head>
|
||||
<title>Enable Developer Mode</title>
|
||||
<style>
|
||||
body {
|
||||
font-family: Arial, sans-serif;
|
||||
display: flex;
|
||||
justify-content: center;
|
||||
align-items: center;
|
||||
height: 100vh;
|
||||
margin: 0;
|
||||
background: #f5f5f5;
|
||||
}
|
||||
.container {
|
||||
background: white;
|
||||
padding: 2rem;
|
||||
border-radius: 8px;
|
||||
box-shadow: 0 2px 10px rgba(0,0,0,0.1);
|
||||
text-align: center;
|
||||
}
|
||||
.success {
|
||||
color: #10b981;
|
||||
font-weight: bold;
|
||||
margin-top: 1rem;
|
||||
}
|
||||
button {
|
||||
background: #3b82f6;
|
||||
color: white;
|
||||
border: none;
|
||||
padding: 0.75rem 1.5rem;
|
||||
border-radius: 6px;
|
||||
cursor: pointer;
|
||||
font-size: 1rem;
|
||||
margin-top: 1rem;
|
||||
}
|
||||
button:hover {
|
||||
background: #2563eb;
|
||||
}
|
||||
</style>
|
||||
</head>
|
||||
<body>
|
||||
<div class="container">
|
||||
<h1>Enable Developer Mode</h1>
|
||||
<p>Click the button below to enable Developer Mode for PunimTag.</p>
|
||||
<button onclick="enableDevMode()">Enable Developer Mode</button>
|
||||
<div id="result"></div>
|
||||
</div>
|
||||
<script>
|
||||
function enableDevMode() {
|
||||
localStorage.setItem('punimtag_developer_mode', 'true');
|
||||
const result = document.getElementById('result');
|
||||
result.innerHTML = '<p class="success">✅ Developer Mode enabled! Redirecting...</p>';
|
||||
setTimeout(() => {
|
||||
window.location.href = '/';
|
||||
}, 1500);
|
||||
}
|
||||
|
||||
// Check if already enabled
|
||||
if (localStorage.getItem('punimtag_developer_mode') === 'true') {
|
||||
document.getElementById('result').innerHTML = '<p class="success">✅ Developer Mode is already enabled!</p>';
|
||||
}
|
||||
</script>
|
||||
</body>
|
||||
</html>
|
||||
|
||||
|
||||
@ -39,11 +39,27 @@ export interface SimilarFaceItem {
|
||||
quality_score: number
|
||||
filename: string
|
||||
pose_mode?: string
|
||||
debug_info?: {
|
||||
encoding_length: number
|
||||
encoding_min: number
|
||||
encoding_max: number
|
||||
encoding_mean: number
|
||||
encoding_std: number
|
||||
encoding_first_10: number[]
|
||||
}
|
||||
}
|
||||
|
||||
export interface SimilarFacesResponse {
|
||||
base_face_id: number
|
||||
items: SimilarFaceItem[]
|
||||
debug_info?: {
|
||||
encoding_length: number
|
||||
encoding_min: number
|
||||
encoding_max: number
|
||||
encoding_mean: number
|
||||
encoding_std: number
|
||||
encoding_first_10: number[]
|
||||
}
|
||||
}
|
||||
|
||||
export interface FaceSimilarityPair {
|
||||
@ -97,6 +113,7 @@ export interface AutoMatchRequest {
|
||||
tolerance: number
|
||||
auto_accept?: boolean
|
||||
auto_accept_threshold?: number
|
||||
use_distance_based_thresholds?: boolean
|
||||
}
|
||||
|
||||
export interface AutoMatchFaceItem {
|
||||
@ -217,11 +234,25 @@ export const facesApi = {
|
||||
})
|
||||
return response.data
|
||||
},
|
||||
getSimilar: async (faceId: number, includeExcluded?: boolean): Promise<SimilarFacesResponse> => {
|
||||
getSimilar: async (faceId: number, includeExcluded?: boolean, debug?: boolean): Promise<SimilarFacesResponse> => {
|
||||
const response = await apiClient.get<SimilarFacesResponse>(`/api/v1/faces/${faceId}/similar`, {
|
||||
params: { include_excluded: includeExcluded || false },
|
||||
params: { include_excluded: includeExcluded || false, debug: debug || false },
|
||||
})
|
||||
return response.data
|
||||
const data = response.data
|
||||
|
||||
// Log debug info to browser console if available
|
||||
if (debug && data.debug_info) {
|
||||
console.log('🔍 Base Face Encoding Debug Info:', data.debug_info)
|
||||
}
|
||||
if (debug && data.items) {
|
||||
data.items.forEach((item, index) => {
|
||||
if (item.debug_info) {
|
||||
console.log(`🔍 Similar Face ${index + 1} (ID: ${item.id}) Encoding Debug Info:`, item.debug_info)
|
||||
}
|
||||
})
|
||||
}
|
||||
|
||||
return data
|
||||
},
|
||||
batchSimilarity: async (request: BatchSimilarityRequest): Promise<BatchSimilarityResponse> => {
|
||||
const response = await apiClient.post<BatchSimilarityResponse>('/api/v1/faces/batch-similarity', request)
|
||||
|
||||
@ -7,7 +7,8 @@ import peopleApi, { Person } from '../api/people'
|
||||
import { apiClient } from '../api/client'
|
||||
import { useDeveloperMode } from '../context/DeveloperModeContext'
|
||||
|
||||
const DEFAULT_TOLERANCE = 0.5
|
||||
const DEFAULT_TOLERANCE = 0.6 // Default for regular auto-match (more lenient)
|
||||
const RUN_AUTO_MATCH_TOLERANCE = 0.5 // Tolerance for Run auto-match button (stricter)
|
||||
|
||||
export default function AutoMatch() {
|
||||
const { isDeveloperMode } = useDeveloperMode()
|
||||
@ -451,9 +452,10 @@ export default function AutoMatch() {
|
||||
setBusy(true)
|
||||
try {
|
||||
const response = await facesApi.autoMatch({
|
||||
tolerance,
|
||||
tolerance: RUN_AUTO_MATCH_TOLERANCE, // Use 0.5 for Run auto-match button (stricter)
|
||||
auto_accept: true,
|
||||
auto_accept_threshold: autoAcceptThreshold
|
||||
auto_accept_threshold: autoAcceptThreshold,
|
||||
use_distance_based_thresholds: true // Enable distance-based thresholds for Run auto-match button
|
||||
})
|
||||
|
||||
// Show summary if auto-accept was performed
|
||||
@ -758,7 +760,7 @@ export default function AutoMatch() {
|
||||
)}
|
||||
</div>
|
||||
<div className="mt-2 text-xs text-gray-600 bg-blue-50 border border-blue-200 rounded p-2">
|
||||
<span className="font-medium">ℹ️ Auto-Match Criteria:</span> Only faces with similarity higher than 70% and picture quality higher than 50% will be auto-matched. Profile faces are excluded for better accuracy.
|
||||
<span className="font-medium">ℹ️ Auto-Match Criteria:</span> Only faces with similarity higher than 85% and picture quality higher than 50% will be auto-matched. Profile faces are excluded for better accuracy.
|
||||
</div>
|
||||
</div>
|
||||
|
||||
|
||||
@ -474,7 +474,7 @@ function AutoMatchPageHelp({ onBack }: { onBack: () => void }) {
|
||||
<li>Click "🚀 Run Auto-Match" button</li>
|
||||
<li>The system will automatically match unidentified faces to identified people based on:
|
||||
<ul className="list-disc list-inside ml-4 mt-1">
|
||||
<li>Similarity higher than 70%</li>
|
||||
<li>Similarity higher than 85%</li>
|
||||
<li>Picture quality higher than 50%</li>
|
||||
<li>Profile faces are excluded for better accuracy</li>
|
||||
</ul>
|
||||
|
||||
@ -348,7 +348,8 @@ export default function Identify() {
|
||||
return
|
||||
}
|
||||
try {
|
||||
const res = await facesApi.getSimilar(faceId, includeExcludedFaces)
|
||||
// Enable debug mode to log encoding info to browser console
|
||||
const res = await facesApi.getSimilar(faceId, includeExcludedFaces, true)
|
||||
setSimilar(res.items || [])
|
||||
setSelectedSimilar({})
|
||||
} catch (error) {
|
||||
|
||||
@ -90,9 +90,9 @@ def process_faces(request: ProcessFacesRequest) -> ProcessFacesResponse:
|
||||
job_timeout="1h", # Long timeout for face processing
|
||||
)
|
||||
|
||||
print(f"[Faces API] Enqueued face processing job: {job.id}")
|
||||
print(f"[Faces API] Job status: {job.get_status()}")
|
||||
print(f"[Faces API] Queue length: {len(queue)}")
|
||||
import logging
|
||||
logger = logging.getLogger(__name__)
|
||||
logger.info(f"Enqueued face processing job: {job.id}, status: {job.get_status()}, queue length: {len(queue)}")
|
||||
|
||||
return ProcessFacesResponse(
|
||||
job_id=job.id,
|
||||
@ -197,12 +197,14 @@ def get_unidentified_faces(
|
||||
def get_similar_faces(
|
||||
face_id: int,
|
||||
include_excluded: bool = Query(False, description="Include excluded faces in results"),
|
||||
debug: bool = Query(False, description="Include debug information (encoding stats) in response"),
|
||||
db: Session = Depends(get_db)
|
||||
) -> SimilarFacesResponse:
|
||||
"""Return similar unidentified faces for a given face."""
|
||||
import logging
|
||||
import numpy as np
|
||||
logger = logging.getLogger(__name__)
|
||||
logger.info(f"API: get_similar_faces called for face_id={face_id}, include_excluded={include_excluded}")
|
||||
logger.info(f"API: get_similar_faces called for face_id={face_id}, include_excluded={include_excluded}, debug={debug}")
|
||||
|
||||
# Validate face exists
|
||||
base = db.query(Face).filter(Face.id == face_id).first()
|
||||
@ -210,9 +212,23 @@ def get_similar_faces(
|
||||
logger.warning(f"API: Face {face_id} not found")
|
||||
raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail=f"Face {face_id} not found")
|
||||
|
||||
# Load base encoding for debug info if needed
|
||||
base_debug_info = None
|
||||
if debug:
|
||||
from backend.services.face_service import load_face_encoding
|
||||
base_enc = load_face_encoding(base.encoding)
|
||||
base_debug_info = {
|
||||
"encoding_length": len(base_enc),
|
||||
"encoding_min": float(np.min(base_enc)),
|
||||
"encoding_max": float(np.max(base_enc)),
|
||||
"encoding_mean": float(np.mean(base_enc)),
|
||||
"encoding_std": float(np.std(base_enc)),
|
||||
"encoding_first_10": [float(x) for x in base_enc[:10].tolist()],
|
||||
}
|
||||
|
||||
logger.info(f"API: Calling find_similar_faces for face_id={face_id}, include_excluded={include_excluded}")
|
||||
# Use 0.6 tolerance for Identify People (more lenient for manual review)
|
||||
results = find_similar_faces(db, face_id, tolerance=0.6, include_excluded=include_excluded)
|
||||
results = find_similar_faces(db, face_id, tolerance=0.6, include_excluded=include_excluded, debug=debug)
|
||||
logger.info(f"API: find_similar_faces returned {len(results)} results")
|
||||
|
||||
items = [
|
||||
@ -224,12 +240,13 @@ def get_similar_faces(
|
||||
quality_score=float(f.quality_score),
|
||||
filename=f.photo.filename if f.photo else "unknown",
|
||||
pose_mode=getattr(f, "pose_mode", None) or "frontal",
|
||||
debug_info=debug_info if debug else None,
|
||||
)
|
||||
for f, distance, confidence_pct in results
|
||||
for f, distance, confidence_pct, debug_info in results
|
||||
]
|
||||
|
||||
logger.info(f"API: Returning {len(items)} items for face_id={face_id}")
|
||||
return SimilarFacesResponse(base_face_id=face_id, items=items)
|
||||
return SimilarFacesResponse(base_face_id=face_id, items=items, debug_info=base_debug_info)
|
||||
|
||||
|
||||
@router.post("/batch-similarity", response_model=BatchSimilarityResponse)
|
||||
@ -438,7 +455,9 @@ def get_face_crop(face_id: int, db: Session = Depends(get_db)) -> Response:
|
||||
except HTTPException:
|
||||
raise
|
||||
except Exception as e:
|
||||
print(f"[Faces API] get_face_crop error for face {face_id}: {e}")
|
||||
import logging
|
||||
logger = logging.getLogger(__name__)
|
||||
logger.error(f"get_face_crop error for face {face_id}: {e}")
|
||||
raise HTTPException(
|
||||
status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
|
||||
detail=f"Failed to extract face crop: {str(e)}",
|
||||
@ -610,10 +629,12 @@ def auto_match_faces(
|
||||
|
||||
# Find matches for all identified people
|
||||
# Filter by frontal reference faces if auto_accept enabled
|
||||
# Use distance-based thresholds only when auto_accept is enabled (Run auto-match button)
|
||||
matches_data = find_auto_match_matches(
|
||||
db,
|
||||
tolerance=request.tolerance,
|
||||
filter_frontal_only=request.auto_accept
|
||||
filter_frontal_only=request.auto_accept,
|
||||
use_distance_based_thresholds=request.use_distance_based_thresholds or request.auto_accept
|
||||
)
|
||||
|
||||
# If auto_accept enabled, process matches automatically
|
||||
@ -647,7 +668,9 @@ def auto_match_faces(
|
||||
)
|
||||
auto_accepted_faces += identified_count
|
||||
except Exception as e:
|
||||
print(f"Error auto-accepting matches for person {person_id}: {e}")
|
||||
import logging
|
||||
logger = logging.getLogger(__name__)
|
||||
logger.error(f"Error auto-accepting matches for person {person_id}: {e}")
|
||||
|
||||
if not matches_data:
|
||||
return AutoMatchResponse(
|
||||
@ -750,7 +773,7 @@ def auto_match_faces(
|
||||
@router.get("/auto-match/people", response_model=AutoMatchPeopleResponse)
|
||||
def get_auto_match_people(
|
||||
filter_frontal_only: bool = Query(False, description="Only include frontal/tilted reference faces"),
|
||||
tolerance: float = Query(0.5, ge=0.0, le=1.0, description="Tolerance threshold"),
|
||||
tolerance: float = Query(0.6, ge=0.0, le=1.0, description="Tolerance threshold (default 0.6 for regular auto-match)"),
|
||||
db: Session = Depends(get_db),
|
||||
) -> AutoMatchPeopleResponse:
|
||||
"""Get list of people for auto-match (without matches) - fast initial load.
|
||||
@ -813,7 +836,7 @@ def get_auto_match_people(
|
||||
@router.get("/auto-match/people/{person_id}/matches", response_model=AutoMatchPersonMatchesResponse)
|
||||
def get_auto_match_person_matches(
|
||||
person_id: int,
|
||||
tolerance: float = Query(0.5, ge=0.0, le=1.0, description="Tolerance threshold"),
|
||||
tolerance: float = Query(0.6, ge=0.0, le=1.0, description="Tolerance threshold (default 0.6 for regular auto-match)"),
|
||||
filter_frontal_only: bool = Query(False, description="Only return frontal/tilted faces"),
|
||||
db: Session = Depends(get_db),
|
||||
) -> AutoMatchPersonMatchesResponse:
|
||||
|
||||
@ -89,6 +89,7 @@ class SimilarFaceItem(BaseModel):
|
||||
quality_score: float
|
||||
filename: str
|
||||
pose_mode: Optional[str] = Field("frontal", description="Pose classification (frontal, profile_left, etc.)")
|
||||
debug_info: Optional[dict] = Field(None, description="Debug information (encoding stats) when debug mode is enabled")
|
||||
|
||||
|
||||
class SimilarFacesResponse(BaseModel):
|
||||
@ -98,6 +99,7 @@ class SimilarFacesResponse(BaseModel):
|
||||
|
||||
base_face_id: int
|
||||
items: list[SimilarFaceItem]
|
||||
debug_info: Optional[dict] = Field(None, description="Debug information (base face encoding stats) when debug mode is enabled")
|
||||
|
||||
|
||||
class BatchSimilarityRequest(BaseModel):
|
||||
@ -215,6 +217,7 @@ class AutoMatchRequest(BaseModel):
|
||||
tolerance: float = Field(0.5, ge=0.0, le=1.0, description="Tolerance threshold (lower = stricter matching)")
|
||||
auto_accept: bool = Field(False, description="Enable automatic acceptance of matching faces")
|
||||
auto_accept_threshold: float = Field(70.0, ge=0.0, le=100.0, description="Similarity threshold for auto-acceptance (0-100%)")
|
||||
use_distance_based_thresholds: bool = Field(False, description="Use distance-based confidence thresholds (stricter for borderline distances)")
|
||||
|
||||
|
||||
class AutoMatchFaceItem(BaseModel):
|
||||
|
||||
@ -631,17 +631,21 @@ def process_photo_faces(
|
||||
if face_width is None:
|
||||
face_width = matched_pose_face.get('face_width')
|
||||
pose_mode = PoseDetector.classify_pose_mode(
|
||||
yaw_angle, pitch_angle, roll_angle, face_width
|
||||
yaw_angle, pitch_angle, roll_angle, face_width, landmarks
|
||||
)
|
||||
else:
|
||||
# Can't calculate yaw, use face_width
|
||||
# Can't calculate yaw, use face_width and landmarks for single-eye detection
|
||||
pose_mode = PoseDetector.classify_pose_mode(
|
||||
yaw_angle, pitch_angle, roll_angle, face_width
|
||||
yaw_angle, pitch_angle, roll_angle, face_width, landmarks
|
||||
)
|
||||
elif face_width is not None:
|
||||
# No landmarks available, use face_width only
|
||||
# Try to get landmarks from matched_pose_face if available
|
||||
landmarks_for_classification = None
|
||||
if matched_pose_face:
|
||||
landmarks_for_classification = matched_pose_face.get('landmarks')
|
||||
pose_mode = PoseDetector.classify_pose_mode(
|
||||
yaw_angle, pitch_angle, roll_angle, face_width
|
||||
yaw_angle, pitch_angle, roll_angle, face_width, landmarks_for_classification
|
||||
)
|
||||
else:
|
||||
# No landmarks and no face_width, use default
|
||||
@ -1746,7 +1750,6 @@ def calculate_cosine_distance(encoding1: np.ndarray, encoding2: np.ndarray) -> f
|
||||
# Normalize encodings (matching desktop exactly)
|
||||
norm1 = np.linalg.norm(enc1)
|
||||
norm2 = np.linalg.norm(enc2)
|
||||
|
||||
if norm1 == 0 or norm2 == 0:
|
||||
return 2.0
|
||||
|
||||
@ -1769,6 +1772,32 @@ def calculate_cosine_distance(encoding1: np.ndarray, encoding2: np.ndarray) -> f
|
||||
return 2.0 # Maximum distance on error
|
||||
|
||||
|
||||
def get_distance_based_min_confidence(distance: float) -> float:
|
||||
"""Get minimum confidence threshold based on distance.
|
||||
|
||||
For borderline distances, require higher confidence to reduce false positives.
|
||||
This is used only when use_distance_based_thresholds=True (e.g., in auto-match).
|
||||
|
||||
Args:
|
||||
distance: Cosine distance between faces (0 = identical, 2 = opposite)
|
||||
|
||||
Returns:
|
||||
Minimum confidence percentage (0-100) required for this distance
|
||||
"""
|
||||
if distance <= 0.15:
|
||||
# Very close matches: standard threshold
|
||||
return 50.0
|
||||
elif distance <= 0.20:
|
||||
# Borderline matches: require higher confidence
|
||||
return 70.0
|
||||
elif distance <= 0.25:
|
||||
# Near threshold: require very high confidence
|
||||
return 85.0
|
||||
else:
|
||||
# Far matches: require extremely high confidence
|
||||
return 95.0
|
||||
|
||||
|
||||
def calculate_adaptive_tolerance(base_tolerance: float, face_quality: float) -> float:
|
||||
"""Calculate adaptive tolerance based on face quality, matching desktop exactly."""
|
||||
# Start with base tolerance
|
||||
@ -1936,7 +1965,9 @@ def find_similar_faces(
|
||||
include_excluded: bool = False, # Include excluded faces in results
|
||||
filter_small_faces: bool = False, # Filter out small faces (for auto-match)
|
||||
min_face_size_ratio: float = 0.005, # Minimum face size ratio (0.5% of image)
|
||||
) -> List[Tuple[Face, float, float]]: # Returns (face, distance, confidence_pct)
|
||||
debug: bool = False, # Include debug information (encoding stats)
|
||||
use_distance_based_thresholds: bool = False, # Use distance-based confidence thresholds (for auto-match)
|
||||
) -> List[Tuple[Face, float, float, dict | None]]: # Returns (face, distance, confidence_pct, debug_info)
|
||||
"""Find similar faces matching desktop logic exactly.
|
||||
|
||||
Desktop flow:
|
||||
@ -1963,6 +1994,7 @@ def find_similar_faces(
|
||||
base: Face = db.query(Face).filter(Face.id == face_id).first()
|
||||
if not base:
|
||||
return []
|
||||
|
||||
|
||||
# Load base encoding - auto-detect dtype (supports both float32 and float64)
|
||||
base_enc = load_face_encoding(base.encoding)
|
||||
@ -1986,11 +2018,24 @@ def find_similar_faces(
|
||||
)
|
||||
|
||||
matches: List[Tuple[Face, float, float]] = []
|
||||
|
||||
for f in all_faces:
|
||||
# Load other encoding - auto-detect dtype (supports both float32 and float64)
|
||||
other_enc = load_face_encoding(f.encoding)
|
||||
other_enc = other_enc.copy() # Make a copy to avoid buffer issues
|
||||
|
||||
# Calculate debug info if requested
|
||||
debug_info = None
|
||||
if debug:
|
||||
debug_info = {
|
||||
"encoding_length": len(other_enc),
|
||||
"encoding_min": float(np.min(other_enc)),
|
||||
"encoding_max": float(np.max(other_enc)),
|
||||
"encoding_mean": float(np.mean(other_enc)),
|
||||
"encoding_std": float(np.std(other_enc)),
|
||||
"encoding_first_10": [float(x) for x in other_enc[:10].tolist()],
|
||||
}
|
||||
|
||||
other_quality = float(f.quality_score) if f.quality_score is not None else 0.5
|
||||
|
||||
# Calculate adaptive tolerance based on both face qualities (matching desktop exactly)
|
||||
@ -2001,7 +2046,6 @@ def find_similar_faces(
|
||||
distance = calculate_cosine_distance(base_enc, other_enc)
|
||||
|
||||
# Filter by distance <= adaptive_tolerance (matching desktop find_similar_faces)
|
||||
# DEBUG: Log tolerance usage for troubleshooting
|
||||
if distance <= adaptive_tolerance:
|
||||
# Get photo info (desktop does this in find_similar_faces)
|
||||
if f.photo:
|
||||
@ -2012,13 +2056,16 @@ def find_similar_faces(
|
||||
# Desktop _get_filtered_similar_faces filters by:
|
||||
# 1. person_id is None (unidentified)
|
||||
# 2. confidence >= 50% (increased from 40% to reduce false matches)
|
||||
# OR confidence >= distance-based threshold if use_distance_based_thresholds=True
|
||||
is_unidentified = f.person_id is None
|
||||
|
||||
# DEBUG: Log all faces that pass distance check
|
||||
if is_unidentified:
|
||||
print(f"DEBUG: Face {f.id} - distance={distance:.4f}, adaptive_tolerance={adaptive_tolerance:.4f}, base_tolerance={tolerance:.4f}, confidence={confidence_pct:.2f}%, passed_distance={distance <= adaptive_tolerance}, passed_confidence={confidence_pct >= 50}")
|
||||
# Calculate minimum confidence threshold
|
||||
if use_distance_based_thresholds:
|
||||
min_confidence = get_distance_based_min_confidence(distance)
|
||||
else:
|
||||
min_confidence = 50.0 # Standard threshold
|
||||
|
||||
if is_unidentified and confidence_pct >= 50:
|
||||
if is_unidentified and confidence_pct >= min_confidence:
|
||||
# Filter by excluded status if not including excluded faces
|
||||
if not include_excluded and getattr(f, "excluded", False):
|
||||
continue
|
||||
@ -2036,7 +2083,7 @@ def find_similar_faces(
|
||||
|
||||
# Return calibrated confidence percentage (matching desktop)
|
||||
# Desktop displays confidence_pct directly from _get_calibrated_confidence
|
||||
matches.append((f, distance, confidence_pct))
|
||||
matches.append((f, distance, confidence_pct, debug_info))
|
||||
|
||||
# Sort by distance (lower is better) - matching desktop
|
||||
matches.sort(key=lambda x: x[1])
|
||||
@ -2197,6 +2244,7 @@ def find_auto_match_matches(
|
||||
db: Session,
|
||||
tolerance: float = 0.5,
|
||||
filter_frontal_only: bool = False,
|
||||
use_distance_based_thresholds: bool = False, # Use distance-based confidence thresholds
|
||||
) -> List[Tuple[int, int, Face, List[Tuple[Face, float, float]]]]:
|
||||
"""Find auto-match matches for all identified people, matching desktop logic exactly.
|
||||
|
||||
@ -2289,27 +2337,30 @@ def find_auto_match_matches(
|
||||
for person_id, reference_face, person_name in person_faces_list:
|
||||
reference_face_id = reference_face.id
|
||||
|
||||
# Check if reference face is too small (exclude from auto-match)
|
||||
reference_photo = db.query(Photo).filter(Photo.id == reference_face.photo_id).first()
|
||||
if reference_photo:
|
||||
ref_size_ratio = _calculate_face_size_ratio(reference_face, reference_photo)
|
||||
if ref_size_ratio < MIN_AUTO_MATCH_FACE_SIZE_RATIO:
|
||||
# Skip this person - reference face is too small
|
||||
continue
|
||||
# TEMPORARILY DISABLED: Check if reference face is too small (exclude from auto-match)
|
||||
# reference_photo = db.query(Photo).filter(Photo.id == reference_face.photo_id).first()
|
||||
# if reference_photo:
|
||||
# ref_size_ratio = _calculate_face_size_ratio(reference_face, reference_photo)
|
||||
# if ref_size_ratio < MIN_AUTO_MATCH_FACE_SIZE_RATIO:
|
||||
# # Skip this person - reference face is too small
|
||||
# continue
|
||||
|
||||
# Use find_similar_faces which matches desktop _get_filtered_similar_faces logic
|
||||
# Desktop: similar_faces = self.face_processor._get_filtered_similar_faces(
|
||||
# reference_face_id, tolerance, include_same_photo=False, face_status=None)
|
||||
# This filters by: person_id is None (unidentified), confidence >= 50% (increased from 40%), sorts by distance
|
||||
# Auto-match always excludes excluded faces
|
||||
# filter_small_faces=True to exclude small match faces
|
||||
similar_faces = find_similar_faces(
|
||||
# TEMPORARILY DISABLED: filter_small_faces=True to exclude small match faces
|
||||
similar_faces_with_debug = find_similar_faces(
|
||||
db, reference_face_id, tolerance=tolerance,
|
||||
filter_frontal_only=filter_frontal_only,
|
||||
include_excluded=False, # Auto-match always excludes excluded faces
|
||||
filter_small_faces=True, # Exclude small faces from auto-match
|
||||
min_face_size_ratio=MIN_AUTO_MATCH_FACE_SIZE_RATIO
|
||||
filter_small_faces=False, # TEMPORARILY DISABLED: Exclude small faces from auto-match
|
||||
min_face_size_ratio=MIN_AUTO_MATCH_FACE_SIZE_RATIO,
|
||||
use_distance_based_thresholds=use_distance_based_thresholds # Use distance-based thresholds if enabled
|
||||
)
|
||||
# Strip debug_info for internal use
|
||||
similar_faces = [(f, dist, conf) for f, dist, conf, _ in similar_faces_with_debug]
|
||||
|
||||
if similar_faces:
|
||||
results.append((person_id, reference_face_id, reference_face, similar_faces))
|
||||
@ -2453,14 +2504,13 @@ def get_auto_match_person_matches(
|
||||
|
||||
# Find similar faces using existing function
|
||||
# Auto-match always excludes excluded faces
|
||||
# DEBUG: Log tolerance being used
|
||||
print(f"DEBUG get_auto_match_person_matches: person_id={person_id}, tolerance={tolerance}, reference_face_id={reference_face.id}")
|
||||
similar_faces = find_similar_faces(
|
||||
similar_faces_with_debug = find_similar_faces(
|
||||
db, reference_face.id, tolerance=tolerance,
|
||||
filter_frontal_only=filter_frontal_only,
|
||||
include_excluded=False # Auto-match always excludes excluded faces
|
||||
)
|
||||
print(f"DEBUG get_auto_match_person_matches: Found {len(similar_faces)} matches for person {person_id} with tolerance {tolerance}")
|
||||
# Strip debug_info for internal use
|
||||
similar_faces = [(f, dist, conf) for f, dist, conf, _ in similar_faces_with_debug]
|
||||
|
||||
return similar_faces
|
||||
|
||||
|
||||
@ -22,7 +22,7 @@ class PoseDetector:
|
||||
"""Detect face pose (yaw, pitch, roll) using RetinaFace landmarks"""
|
||||
|
||||
# Thresholds for pose detection (in degrees)
|
||||
PROFILE_YAW_THRESHOLD = 30.0 # Faces with |yaw| >= 30° are considered profile
|
||||
PROFILE_YAW_THRESHOLD = 15.0 # Faces with |yaw| >= 15° are considered profile
|
||||
EXTREME_YAW_THRESHOLD = 60.0 # Faces with |yaw| >= 60° are extreme profile
|
||||
|
||||
PITCH_THRESHOLD = 20.0 # Faces with |pitch| >= 20° are looking up/down
|
||||
@ -39,7 +39,7 @@ class PoseDetector:
|
||||
|
||||
Args:
|
||||
yaw_threshold: Yaw angle threshold for profile detection (degrees)
|
||||
Default: 30.0
|
||||
Default: 15.0
|
||||
pitch_threshold: Pitch angle threshold for up/down detection (degrees)
|
||||
Default: 20.0
|
||||
roll_threshold: Roll angle threshold for tilt detection (degrees)
|
||||
@ -53,17 +53,24 @@ class PoseDetector:
|
||||
self.roll_threshold = roll_threshold or self.ROLL_THRESHOLD
|
||||
|
||||
@staticmethod
|
||||
def detect_faces_with_landmarks(img_path: str) -> Dict:
|
||||
def detect_faces_with_landmarks(img_path: str, filter_estimated_landmarks: bool = False) -> Dict:
|
||||
"""Detect faces using RetinaFace directly
|
||||
|
||||
Args:
|
||||
img_path: Path to image file
|
||||
filter_estimated_landmarks: If True, remove landmarks that appear to be estimated
|
||||
(e.g., hidden eye in profile views) rather than actually visible.
|
||||
Uses heuristics: if eyes are very close together (< 20px) and
|
||||
yaw calculation suggests extreme profile, mark hidden eye as None.
|
||||
|
||||
Returns:
|
||||
Dictionary with face keys and landmark data:
|
||||
{
|
||||
'face_1': {
|
||||
'facial_area': {'x': x, 'y': y, 'w': w, 'h': h},
|
||||
'landmarks': {
|
||||
'left_eye': (x, y),
|
||||
'right_eye': (x, y),
|
||||
'left_eye': (x, y) or None,
|
||||
'right_eye': (x, y) or None,
|
||||
'nose': (x, y),
|
||||
'left_mouth': (x, y),
|
||||
'right_mouth': (x, y)
|
||||
@ -76,6 +83,42 @@ class PoseDetector:
|
||||
return {}
|
||||
|
||||
faces = RetinaFace.detect_faces(img_path)
|
||||
|
||||
# Post-process to filter estimated landmarks if requested
|
||||
if filter_estimated_landmarks:
|
||||
for face_key, face_data in faces.items():
|
||||
landmarks = face_data.get('landmarks', {})
|
||||
if not landmarks:
|
||||
continue
|
||||
|
||||
left_eye = landmarks.get('left_eye')
|
||||
right_eye = landmarks.get('right_eye')
|
||||
nose = landmarks.get('nose')
|
||||
|
||||
# Check if both eyes are present and very close together (profile view)
|
||||
if left_eye and right_eye and nose:
|
||||
face_width = abs(right_eye[0] - left_eye[0])
|
||||
|
||||
# If eyes are very close (< 20px), likely a profile view
|
||||
if face_width < 20.0:
|
||||
# Calculate which eye is likely hidden based on nose position
|
||||
eye_mid_x = (left_eye[0] + right_eye[0]) / 2
|
||||
nose_x = nose[0]
|
||||
|
||||
# If nose is closer to left eye, right eye is likely hidden (face turned left)
|
||||
# If nose is closer to right eye, left eye is likely hidden (face turned right)
|
||||
dist_to_left = abs(nose_x - left_eye[0])
|
||||
dist_to_right = abs(nose_x - right_eye[0])
|
||||
|
||||
if dist_to_left < dist_to_right:
|
||||
# Nose closer to left eye = face turned left = right eye hidden
|
||||
landmarks['right_eye'] = None
|
||||
else:
|
||||
# Nose closer to right eye = face turned right = left eye hidden
|
||||
landmarks['left_eye'] = None
|
||||
|
||||
face_data['landmarks'] = landmarks
|
||||
|
||||
return faces
|
||||
|
||||
@staticmethod
|
||||
@ -260,7 +303,8 @@ class PoseDetector:
|
||||
def classify_pose_mode(yaw: Optional[float],
|
||||
pitch: Optional[float],
|
||||
roll: Optional[float],
|
||||
face_width: Optional[float] = None) -> str:
|
||||
face_width: Optional[float] = None,
|
||||
landmarks: Optional[Dict] = None) -> str:
|
||||
"""Classify face pose mode from all three angles and optionally face width
|
||||
|
||||
Args:
|
||||
@ -268,8 +312,10 @@ class PoseDetector:
|
||||
pitch: Pitch angle in degrees
|
||||
roll: Roll angle in degrees
|
||||
face_width: Face width in pixels (eye distance). Used as indicator for profile detection.
|
||||
If face_width < 25px, indicates profile view. When yaw is available but < 30°,
|
||||
If face_width < 25px, indicates profile view. When yaw is available but < 15°,
|
||||
face_width can override yaw if it suggests profile (face_width < 25px).
|
||||
landmarks: Optional facial landmarks dictionary. Used to detect single-eye visibility
|
||||
for extreme profile views where only one eye is visible.
|
||||
|
||||
Returns:
|
||||
Pose mode classification string:
|
||||
@ -279,6 +325,28 @@ class PoseDetector:
|
||||
- 'tilted_left', 'tilted_right': roll variations
|
||||
- Combined modes: e.g., 'profile_left_looking_up'
|
||||
"""
|
||||
# Check for single-eye visibility to infer profile direction
|
||||
# This handles extreme profile views where only one eye is visible
|
||||
if landmarks:
|
||||
left_eye = landmarks.get('left_eye')
|
||||
right_eye = landmarks.get('right_eye')
|
||||
|
||||
# Only right eye visible -> face turned left -> profile_left
|
||||
if left_eye is None and right_eye is not None:
|
||||
# Infer profile_left when only right eye is visible
|
||||
inferred_profile = "profile_left"
|
||||
# Only left eye visible -> face turned right -> profile_right
|
||||
elif left_eye is not None and right_eye is None:
|
||||
# Infer profile_right when only left eye is visible
|
||||
inferred_profile = "profile_right"
|
||||
# No eyes visible -> extreme profile, default to profile_left
|
||||
elif left_eye is None and right_eye is None:
|
||||
inferred_profile = "profile_left"
|
||||
else:
|
||||
inferred_profile = None # Both eyes visible, use normal logic
|
||||
else:
|
||||
inferred_profile = None
|
||||
|
||||
# Default to frontal if angles unknown
|
||||
yaw_original = yaw
|
||||
if yaw is None:
|
||||
@ -290,20 +358,23 @@ class PoseDetector:
|
||||
|
||||
# Face width threshold for profile detection (in pixels)
|
||||
# Profile faces have very small eye distance (< 25 pixels typically)
|
||||
PROFILE_FACE_WIDTH_THRESHOLD = 10.0 #25.0
|
||||
PROFILE_FACE_WIDTH_THRESHOLD = 20.0
|
||||
|
||||
# Yaw classification - PRIMARY INDICATOR
|
||||
# Use yaw angle as the primary indicator (30° threshold)
|
||||
# Use yaw angle as the primary indicator (15° threshold)
|
||||
abs_yaw = abs(yaw)
|
||||
|
||||
# Primary classification based on yaw angle
|
||||
if abs_yaw < 30.0:
|
||||
if abs_yaw < 15.0:
|
||||
# Yaw indicates frontal view
|
||||
# Trust yaw when it's available and reasonable (< 30°)
|
||||
# Trust yaw when it's available and reasonable (< 15°)
|
||||
# Only use face_width as fallback when yaw is unavailable (None)
|
||||
if yaw_original is None:
|
||||
# Yaw unavailable - use face_width as fallback
|
||||
if face_width is not None:
|
||||
# Yaw unavailable - check for single-eye visibility first
|
||||
if inferred_profile is not None:
|
||||
# Single eye visible or no eyes visible -> use inferred profile direction
|
||||
yaw_mode = inferred_profile
|
||||
elif face_width is not None:
|
||||
if face_width < PROFILE_FACE_WIDTH_THRESHOLD:
|
||||
# Face width suggests profile view - use it when yaw is unavailable
|
||||
yaw_mode = "profile_left" # Default direction when yaw unavailable
|
||||
@ -311,16 +382,14 @@ class PoseDetector:
|
||||
# Face width is normal (>= 25px) - likely frontal
|
||||
yaw_mode = "frontal"
|
||||
else:
|
||||
# Both yaw and face_width unavailable - cannot determine reliably
|
||||
# This usually means landmarks are incomplete (missing nose and/or eyes)
|
||||
# For extreme profile views, both eyes might not be visible, which would
|
||||
# cause face_width to be None. In this case, we cannot reliably determine
|
||||
# pose without additional indicators (like face bounding box aspect ratio).
|
||||
# Default to frontal (conservative approach), but this might misclassify
|
||||
# some extreme profile faces.
|
||||
yaw_mode = "frontal"
|
||||
# Both yaw and face_width unavailable - check if we inferred profile from landmarks
|
||||
if inferred_profile is not None:
|
||||
yaw_mode = inferred_profile
|
||||
else:
|
||||
# Cannot determine reliably - default to frontal
|
||||
yaw_mode = "frontal"
|
||||
else:
|
||||
# Yaw is available and < 30° - but still check face_width
|
||||
# Yaw is available and < 15° - but still check face_width
|
||||
# If face_width is very small (< 25px), it suggests profile even with small yaw
|
||||
if face_width is not None:
|
||||
if face_width < PROFILE_FACE_WIDTH_THRESHOLD:
|
||||
@ -332,11 +401,11 @@ class PoseDetector:
|
||||
else:
|
||||
# No face_width provided - trust yaw, classify as frontal
|
||||
yaw_mode = "frontal"
|
||||
elif yaw <= -30.0:
|
||||
# abs_yaw >= 30.0 and yaw is negative - profile left
|
||||
elif yaw <= -15.0:
|
||||
# abs_yaw >= 15.0 and yaw is negative - profile left
|
||||
yaw_mode = "profile_left" # Negative yaw = face turned left = left profile visible
|
||||
elif yaw >= 30.0:
|
||||
# abs_yaw >= 30.0 and yaw is positive - profile right
|
||||
elif yaw >= 15.0:
|
||||
# abs_yaw >= 15.0 and yaw is positive - profile right
|
||||
yaw_mode = "profile_right" # Positive yaw = face turned right = right profile visible
|
||||
else:
|
||||
# This should never be reached, but handle edge case
|
||||
@ -411,8 +480,8 @@ class PoseDetector:
|
||||
# Calculate face width (eye distance) for profile detection
|
||||
face_width = self.calculate_face_width_from_landmarks(landmarks)
|
||||
|
||||
# Classify pose mode (using face width as additional indicator)
|
||||
pose_mode = self.classify_pose_mode(yaw_angle, pitch_angle, roll_angle, face_width)
|
||||
# Classify pose mode (using face width and landmarks as additional indicators)
|
||||
pose_mode = self.classify_pose_mode(yaw_angle, pitch_angle, roll_angle, face_width, landmarks)
|
||||
|
||||
# Normalize facial_area format (RetinaFace returns list [x, y, w, h] or dict)
|
||||
facial_area_raw = face_data.get('facial_area', {})
|
||||
|
||||
@ -111,3 +111,4 @@ In CI (GitHub Actions/Gitea Actions), test results appear in:
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
@ -207,3 +207,4 @@ echo ""
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
@ -148,3 +148,4 @@ testQueries()
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
@ -25,3 +25,4 @@ fi
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user