Its working pretty well now

Signed-off-by: James Ketrenos <james_git@ketrenos.com>

Dockerfile

@@ -20,7 +20,8 @@ RUN wget -qO- https://deb.nodesource.com/setup_18.x | bash -
 RUN DEBIAN_FRONTEND=NONINTERACTIVE apt-get install -y \
   python2 \
   jhead \
-  nodejs
+  nodejs \
+  jq
 
 # Install the latest npm and npx
 RUN npm install --global npm@latest

docker-compose.yml

@@ -9,9 +9,9 @@ services:
  #     - db
     restart: always
     ports:
-      - 8134:8123
+      - 8135:8123
     volumes:
-      - /multimedia/Dad:/pictures
+      - /home/jketreno/.pic-chalkwharf-bk:/pictures
       - ${PWD}/db:/db
       - ${PWD}:/website
       - ${PWD}/models:/root/.deepface

ketrface/cluster.py

@@ -6,8 +6,7 @@ import sqlite3
 from sqlite3 import Error
 from PIL import Image
 import numpy as np
-from deepface import DeepFace
-from deepface.detectors import FaceDetector
+
 import functools
 
 from ketrface.util import *
@@ -41,6 +40,7 @@ def gen_html(identities):
       photoId = face['photoId']
       distance = "{:0.4f}".format(face['distance'])
       confidence = "{:0.3f}".format(face['confidence'])
+      focus = int(face['focus'])
       label = face['cluster']
       if type(label) != str:
         label = f'Cluster ({face["cluster"]["id"]})'
@@ -49,25 +49,40 @@ def gen_html(identities):
       path = f'{html_base}/faces/{"{:02d}".format(faceId % 10)}'
       print(f'<img src="{path}/{faceId}.jpg"/>')
       print(f'<div style="background-color:rgba(255, 255, 255, 0.4);position:absolute;top:0px;left:0px;right:0px;padding:0.25rem">{label}: {distance}</div>')
-      print(f'<div style="background-color:rgba(255, 255, 255, 0.4);position:absolute;bottom:0px;left:0px;right:0px;padding:0.25rem">{faceId} {photoId} {confidence}</div>')
+      print(f'<div style="background-color:rgba(255, 255, 255, 0.4);position:absolute;bottom:0px;left:0px;right:0px;padding:0.25rem">{faceId} {photoId} {confidence} {focus}</div>')
       print('</div>')
     print('</div>')
     print('</div>')
 
+def update_cluster_averages(identities):
+  for identity in identities:
+    average = []
+    for face in identity['faces']:
+      if len(average) == 0:
+        average = face['descriptors']
+      else:
+        average = np.add(average, face['descriptors'])
+    average = np.divide(average, len(identity['faces']))
+    identity['descriptors'] = average
+  return identities
 
 def load_faces(db_path = db_path):
+  print(f'Connecting to database: {db_path}')
   conn = create_connection(db_path)
   faces = []
   with conn:
+    print('Querying faces')
     cur = conn.cursor()
     res = cur.execute('''
-      SELECT faces.id,facedescriptors.descriptors,faces.faceConfidence,faces.photoId
+      SELECT faces.id,facedescriptors.descriptors,faces.faceConfidence,faces.photoId,faces.focus
       FROM faces
       JOIN facedescriptors ON (faces.descriptorId=facedescriptors.id)
       WHERE faces.identityId IS null AND faces.faceConfidence>0.99
       ''')
     for row in res.fetchall():
-      id, descriptors, confidence, photoId = row
+      id, descriptors, confidence, photoId, focus = row
+      if focus is None:
+        focus = 100 # Assume full focus if focus not set
       face = { 
         'id': id,
         'type': 'face',
@@ -75,12 +90,38 @@ def load_faces(db_path = db_path):
         'distance': 0,
         'photoId': photoId,
         'descriptors': np.frombuffer(descriptors),
-        'cluster': Undefined
+        'cluster': Undefined,
+        'focus': focus
       }
       face['faces'] = [ face ]
       faces.append(face)
   return faces
 
+def update_distances(identities, prune = False):
+  removed = 0
+  for identity in identities:
+    for face in identity['faces']:
+      average = identity['descriptors']
+      distance = findCosineDistance(average, face['descriptors'])
+      if prune and distance > MAX_EPOCH_DISTANCE:
+        average = np.dot(average, len(identity['faces']))
+        average = np.subtract(average, face['descriptors'])
+  
+        face['cluster'] = Undefined
+        face['distance'] = 0
+        identity['faces'].remove(face)
+
+        identity['descriptors'] = np.divide(average, len(identity['faces']))
+        removed += 1
+      else:
+        face['distance'] = distance
+  return removed
+
+def sort_identities(identities):
+  identities.sort(reverse = True, key = lambda x: len(x['faces']))
+  for identity in identities:
+    identity['faces'].sort(reverse = False, key = lambda x: x['distance'])
+
 def cluster_sort(A, B):
   diff = A['cluster'] - B['cluster']
   if diff > 0:
@@ -101,23 +142,11 @@ print('Scanning for clusters')
 identities = DBSCAN(faces) # process_faces(faces)
 print(f'{len(identities)} clusters grouped')
 
-
+MAX_CLUSTER_DISTANCE = 0.15  # Used to merge clusters
+MAX_EPOCH_DISTANCE = 0.14    # Used to prune outliers
 
 # Compute average center for all clusters
-sum = 0
-for identity in identities:
-  sum += len(identity['faces'])
-  print(f'{identity["id"]} has {len(identity["faces"])} faces')
-  average = []
-
-  for face in identity['faces']:
-    if len(average) == 0:
-      average = face['descriptors']
-    else:
-      average = np.add(average, face['descriptors'])
-
-  average = np.divide(average, len(identity['faces']))
-  identity['descriptors'] = average
+identities = update_cluster_averages(identities)
 
 removed = -1
 epoch = 1
@@ -126,33 +155,68 @@ epoch = 1
 while removed != 0:
   print(f'Epoch {epoch}...')
   epoch += 1
-  removed = 0
-  for identity in identities:
-    for face in identity['faces']:
-      average = identity['descriptors']
-      distance = findCosineDistance(average, face['descriptors'])
-      if distance > 0.14:
-        average = np.dot(average, len(identity['faces']))
-        average = np.subtract(average, face['descriptors'])
-  
-        face['cluster'] = Undefined
-        face['distance'] = 0
-        identity['faces'].remove(face)
-
-        identity['descriptors'] = np.divide(average, len(identity['faces']))
-        removed += 1
-      else:
-        face['distance'] = distance
+  removed = update_distances(identities, prune = True)
   if removed > 0:
     print(f'Excluded {removed} faces this epoch')
-    
-identities.sort(reverse = True, key = lambda x: len(x['faces']))
-for identity in identities:
-  identity['faces'].sort(reverse = False, key = lambda x: x['distance'])
+
+
 
 print(f'{len(identities)} identities seeded.')
 
+# Cluster the clusters...
+print('Reducing clusters via DBSCAN')
+reduced = DBSCAN(identities, eps = MAX_CLUSTER_DISTANCE, minPts = 2)
+# For each cluster, merge the lists of faces referenced in the cluster's
+# "faces" field, which is pointing to clusters (and not actual faces)
+for cluster in reduced:
+  merged = []
+  for identity in cluster['faces']:
+    merged = merged + identity['faces']
+  cluster['faces'] = merged
+
+# Creating a set containing those faces which have not been bound
+# to an identity to recluster them in isolation from the rest of
+# the faces
+noise = []
+undefined = []
+clustered = []
+for face in faces:
+  if face['cluster'] == Noise:
+    noise.append(face)
+  elif face['cluster'] == Undefined:
+    undefined.append(face)
+
+print(f'Stats: Noise = {len(noise)}, Undefined = {len(undefined)}')
+
+straglers = DBSCAN(noise + undefined)
+reduced = update_cluster_averages(reduced + straglers)
+
+# Give all merged identity lists a unique ID
+for id, identity in enumerate(reduced):
+  identity['id'] = id
+  for face in identity['faces']:
+    face['cluster'] = identity
+
+update_distances(reduced)
+
+sort_identities(reduced)
+
+# This generates a set of differences between clusters and makes
+# a recommendation to merge clusters (outside of DBSCAN)
+#
+# Worth testing on larger data set
+for i, A in enumerate(reduced):
+  for k, B in enumerate(reduced):
+    if k < i:
+      continue
+    if A == B:
+      continue
+    distance = findCosineDistance(A['descriptors'], B['descriptors'])
+    if distance < MAX_CLUSTER_DISTANCE:
+      distance = "{:0.4f}".format(distance)
+      print(f'{A["id"]} to {B["id"]} = {distance}: MERGE')
+
 print('Writing to "identities.html"')
 redirect_on('identities.html')
-gen_html(identities)
+gen_html(reduced)
 redirect_off()

ketrface/detect.py

@@ -1,5 +1,4 @@
 import sys
-import zlib
 import json
 import os
 import piexif
@@ -13,6 +12,7 @@ import cv2
 from ketrface.util import *
 from ketrface.db import *
 
+face_base = '../'
 model_name = 'VGG-Face'    # 'ArcFace'
 detector_backend = 'mtcnn' # 'retinaface'
 model = DeepFace.build_model(model_name)
@@ -92,18 +92,22 @@ def extract_faces(img, threshold=0.95, allow_upscaling = True, focus_threshold =
             'right_eye': list(face['keypoints']['right_eye']),
           },
           'score': face['confidence'],
+          
         }
 
+  to_drop = []
+
   # Re-implementation of 'extract_faces' with the addition of keeping a 
   # copy of the face image for caching on disk
   for k, key in enumerate(faces):
     print(f'Processing face {k+1}/{len(faces)}')
     identity = faces[key]
+    identity['focus'] = 100 # Until laplacian variance is executed
+
     facial_area = identity["facial_area"]
     landmarks = identity["landmarks"]
     left_eye = landmarks["left_eye"]
     right_eye = landmarks["right_eye"]
-
 #      markup = True
     markup = False
     if markup == True: # Draw the face rectangle and eyes
@@ -142,16 +146,12 @@ def extract_faces(img, threshold=0.95, allow_upscaling = True, focus_threshold =
 
     facial_img = img[top: bottom, left: right]
 
-    image = Image.fromarray(facial_img)
-
-    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+    gray = cv2.cvtColor(facial_img, cv2.COLOR_BGR2GRAY)
     focus = variance_of_laplacian(gray)
-    if focus < focus_threshold:
-      print(f'Dropping {ke+1} due to focus {focus}.')
-      faces.pop(key)
     identity['focus'] = focus
 
     # Eye order is reversed as the routine does them backwards
+    image = Image.fromarray(facial_img)
     image = alignment_procedure(image, right_eye, left_eye)
     image = image.resize(size = input_shape, resample = Image.LANCZOS)
     resized = np.asarray(image)
@@ -174,6 +174,9 @@ def extract_faces(img, threshold=0.95, allow_upscaling = True, focus_threshold =
 
     identity['image'] = Image.fromarray(resized)
 
+#  for key in to_drop:
+#    faces.pop(key)
+ 
   return faces
 
 
@@ -192,9 +195,14 @@ with conn:
     photoId, photoFaces, albumPath, photoFilename = row
     img_path = f'{base}{albumPath}{photoFilename}'
     print(f'Processing {i+1}/{count}: {img_path}')
-    img = Image.open(img_path)
-    img = ImageOps.exif_transpose(img) # auto-rotate if needed
-    img = img.convert()
+    try:
+      img = Image.open(img_path)
+      img = ImageOps.exif_transpose(img) # auto-rotate if needed
+      img = img.convert("RGB") # Catch "RGBA" and convert to 3-channel
+    except:
+      print(f'Unable to load / process {img_path}. Skipping.')
+      continue
+
     img = np.asarray(img)
     faces = extract_faces(img)
     if faces is None:
@@ -221,6 +229,7 @@ with conn:
         'photoId': photoId,
         'scanVersion': face['version'],
         'faceConfidence': face['score'],
+        'focus': face['focus'],
         'top': face['face']['top'],
         'left': face['face']['left'],
         'bottom': face['face']['bottom'],
@@ -228,7 +237,7 @@ with conn:
         'descriptorId': faceDescriptorId,
       })
 
-      path = f'faces/{"{:02d}".format(faceId % 10)}'
+      path = f'{face_base}faces/{"{:02d}".format(faceId % 10)}'
       try:
         os.mkdir(path)
       except FileExistsError:

ketrface/headers.py

@@ -8,13 +8,14 @@ from io import BytesIO
 
 from ketrface.util import *
 
+face_base = "../"
 faceId = int(sys.argv[1])
-path = f'faces/{"{:02d}".format(faceId % 10)}'
+path = f'{face_base}faces/{"{:02d}".format(faceId % 10)}'
 
 img = Image.open(f'{path}/{faceId}.jpg')
 exif_dict = piexif.load(img.info["exif"])
 compressed_str = exif_dict["Exif"][piexif.ExifIFD.UserComment]
 
 str = zlib_uudecode(compressed_str)
-json = json.loads(str)
-print(json)
+parsed = json.loads(str)
+print(json.dumps(parsed, indent=2))

ketrface/ketrface/db.py

@@ -27,14 +27,15 @@ def create_face(conn, face):
   :return: face id
   """
   sql = '''
-  INSERT INTO faces(photoId,scanVersion,faceConfidence,top,left,bottom,right,descriptorId)
-  VALUES(?,?,?,?,?,?,?,?)
+  INSERT INTO faces(photoId,scanVersion,faceConfidence,focus,top,left,bottom,right,descriptorId)
+  VALUES(?,?,?,?,?,?,?,?,?)
   '''
   cur = conn.cursor()
   cur.execute(sql, (
     face['photoId'], 
     face['scanVersion'], 
     face['faceConfidence'],
+    face['focus'],
     face['top'],
     face['left'],
     face['bottom'],

ketrface/ketrface/dbscan.py

@@ -1,6 +1,6 @@
 from ketrface.util import *
 
-MIN_PTS = 10
+MIN_PTS = 5
 MAX_DISTANCE = 0.25
 
 Undefined = 0
@@ -25,7 +25,7 @@ def DBSCAN(points, eps = MAX_DISTANCE, minPts = MIN_PTS, verbose = True):
       new_perc = int(100 * (i+1) / total)
       if new_perc != perc:
         perc = new_perc
-        print(f'Clustering faces {perc}% ({i}/{total} processed) complete with {len(clusters)} identities.')
+        print(f'Clustering points {perc}% ({i}/{total} processed) complete with {len(clusters)} identities.')
 
     if P['cluster'] != Undefined:     # Previously processed in inner loop
       continue
@@ -36,7 +36,8 @@ def DBSCAN(points, eps = MAX_DISTANCE, minPts = MIN_PTS, verbose = True):
 
     C = {                             # Define new cluster
       'id': len(clusters),
-      'faces': [ P ]
+      'faces': [ P ],
+      'cluster': Undefined
     }
     clusters.append(C)
 

ketrface/ketrface/util.py

@@ -4,6 +4,7 @@ import uu
 from io import BytesIO
 import json
 import numpy as np
+import zlib
 
 original = None
 

server/db/photos.js

@@ -141,7 +141,11 @@ function init() {
          */
         defaultValue: 0
       },
-
+      focus: { /* focal via laplacian variance https://pyimagesearch.com/2015/09/07/blur-detection-with-opencv/ */
+        type: Sequelize.DOUBLE,
+        defaultValue: 0
+      },
+      
       top: Sequelize.FLOAT, /* 0..1 * photoId.height */
       left: Sequelize.FLOAT, /* 0..1 * photoId.width */
       bottom: Sequelize.FLOAT, /* 0..1 * photoId.height */