Merge pull request #1351 from kremnik/recognition_find_batched

Recognition find batched

deepface/DeepFace.py

@@ -276,7 +276,8 @@ def find(
     silent: bool = False,
     refresh_database: bool = True,
     anti_spoofing: bool = False,
-) -> List[pd.DataFrame]:
+    batched: bool = False,
+) -> Union[List[pd.DataFrame], List[List[Dict[str, Any]]]]:
     """
     Identify individuals in a database
     Args:
@@ -322,22 +323,32 @@ def find(
         anti_spoofing (boolean): Flag to enable anti spoofing (default is False).
 
     Returns:
-        results (List[pd.DataFrame]): A list of pandas dataframes. Each dataframe corresponds
-            to the identity information for an individual detected in the source image.
-            The DataFrame columns include:
+        results (List[pd.DataFrame] or List[List[Dict[str, Any]]]): 
+            A list of pandas dataframes (if `batched=False`) or
+            a list of dicts (if `batched=True`).
+            Each dataframe or dict corresponds to the identity information for
+            an individual detected in the source image.
 
-        - 'identity': Identity label of the detected individual.
+            Note: If you have a large database and/or a source photo with many faces,
+            use `batched=True`, as it is optimized for large batch processing. 
+            Please pay attention that when using `batched=True`, the function returns 
+            a list of dicts (not a list of DataFrames),
+            but with the same keys as the columns in the DataFrame.
+            
+            The DataFrame columns or dict keys include:
 
-        - 'target_x', 'target_y', 'target_w', 'target_h': Bounding box coordinates of the
-                target face in the database.
+            - 'identity': Identity label of the detected individual.
 
-        - 'source_x', 'source_y', 'source_w', 'source_h': Bounding box coordinates of the
-                detected face in the source image.
+            - 'target_x', 'target_y', 'target_w', 'target_h': Bounding box coordinates of the
+                    target face in the database.
 
-        - 'threshold': threshold to determine a pair whether same person or different persons
+            - 'source_x', 'source_y', 'source_w', 'source_h': Bounding box coordinates of the
+                    detected face in the source image.
 
-        - 'distance': Similarity score between the faces based on the
-                specified model and distance metric
+            - 'threshold': threshold to determine a pair whether same person or different persons
+
+            - 'distance': Similarity score between the faces based on the
+                    specified model and distance metric
     """
     return recognition.find(
         img_path=img_path,
@@ -353,6 +364,7 @@ def find(
         silent=silent,
         refresh_database=refresh_database,
         anti_spoofing=anti_spoofing,
+        batched=batched
     )
 
 

deepface/modules/recognition.py

@@ -31,7 +31,8 @@ def find(
     silent: bool = False,
     refresh_database: bool = True,
     anti_spoofing: bool = False,
-) -> List[pd.DataFrame]:
+    batched: bool = False,
+) -> Union[List[pd.DataFrame], List[List[Dict[str, Any]]]]:
     """
     Identify individuals in a database
 
@@ -77,9 +78,19 @@ def find(
 
 
     Returns:
-        results (List[pd.DataFrame]): A list of pandas dataframes. Each dataframe corresponds
-            to the identity information for an individual detected in the source image.
-            The DataFrame columns include:
+        results (List[pd.DataFrame] or List[List[Dict[str, Any]]]): 
+            A list of pandas dataframes (if `batched=False`) or
+            a list of dicts (if `batched=True`).
+            Each dataframe or dict corresponds to the identity information for
+            an individual detected in the source image.
+
+            Note: If you have a large database and/or a source photo with many faces,
+            use `batched=True`, as it is optimized for large batch processing. 
+            Please pay attention that when using `batched=True`, the function returns 
+            a list of dicts (not a list of DataFrames),
+            but with the same keys as the columns in the DataFrame.
+            
+            The DataFrame columns or dict keys include:
 
             - 'identity': Identity label of the detected individual.
 
@@ -233,10 +244,6 @@ def find(
 
     # ----------------------------
     # now, we got representations for facial database
-    df = pd.DataFrame(representations)
-
-    if silent is False:
-        logger.info(f"Searching {img_path} in {df.shape[0]} length datastore")
 
     # img path might have more than once face
     source_objs = detection.extract_faces(
@@ -249,6 +256,24 @@ def find(
         anti_spoofing=anti_spoofing,
     )
 
+    if batched:
+        return find_batched(
+            representations,
+            source_objs,
+            model_name,
+            distance_metric,
+            enforce_detection,
+            align,
+            threshold,
+            normalization,
+            anti_spoofing
+        )
+
+    df = pd.DataFrame(representations)
+
+    if silent is False:
+        logger.info(f"Searching {img_path} in {df.shape[0]} length datastore")
+
     resp_obj = []
 
     for source_obj in source_objs:
@@ -415,3 +440,219 @@ def __find_bulk_embeddings(
                 )
 
     return representations
+
+def find_batched(
+    representations: List[Dict[str, Any]],
+    source_objs: List[Dict[str, Any]],
+    model_name: str = "VGG-Face",
+    distance_metric: str = "cosine",
+    enforce_detection: bool = True,
+    align: bool = True,
+    threshold: Optional[float] = None,
+    normalization: str = "base",
+    anti_spoofing: bool = False,
+) -> List[List[Dict[str, Any]]]:
+    """
+    Perform batched face recognition by comparing source face embeddings with a set of
+    target embeddings. It calculates pairwise distances between the source and target
+    embeddings using the specified distance metric.
+    The function uses batch processing for efficient computation of distances.
+
+    Args:
+        representations (List[Dict[str, Any]]): 
+            A list of dictionaries containing precomputed target embeddings and associated metadata. 
+            Each dictionary should have at least the key `embedding`.
+        
+        source_objs (List[Dict[str, Any]]): 
+            A list of dictionaries representing the source images to compare against
+            the target embeddings. Each dictionary should contain:
+                - `face`: The image data or path to the source face image.
+                - `facial_area`: A dictionary with keys `x`, `y`, `w`, `h`
+                   indicating the facial region.
+                - Optionally, `is_real`: A boolean indicating if the face is real
+                  (used for anti-spoofing).
+        
+        model_name (str): Model for face recognition. Options: VGG-Face, Facenet, Facenet512,
+            OpenFace, DeepFace, DeepID, Dlib, ArcFace, SFace and GhostFaceNet (default is VGG-Face).
+
+        distance_metric (string): Metric for measuring similarity. Options: 'cosine',
+            'euclidean', 'euclidean_l2'.
+
+        enforce_detection (boolean): If no face is detected in an image, raise an exception.
+            Default is True. Set to False to avoid the exception for low-resolution images.
+
+        detector_backend (string): face detector backend. Options: 'opencv', 'retinaface',
+            'mtcnn', 'ssd', 'dlib', 'mediapipe', 'yolov8', 'centerface' or 'skip'.
+
+        align (boolean): Perform alignment based on the eye positions.
+
+        threshold (float): Specify a threshold to determine whether a pair represents the same
+            person or different individuals. This threshold is used for comparing distances.
+            If left unset, default pre-tuned threshold values will be applied based on the specified
+            model name and distance metric (default is None).
+
+        normalization (string): Normalize the input image before feeding it to the model.
+            Default is base. Options: base, raw, Facenet, Facenet2018, VGGFace, VGGFace2, ArcFace
+
+        silent (boolean): Suppress or allow some log messages for a quieter analysis process.
+
+        anti_spoofing (boolean): Flag to enable anti spoofing (default is False).
+
+    Returns:
+        List[List[Dict[str, Any]]]: 
+            A list where each element corresponds to a source face and
+            contains a list of dictionaries with matching faces.
+    """
+    embeddings_list = []
+    valid_mask = []
+    metadata = set()
+
+    for item in representations:
+        emb = item.get('embedding')
+        if emb is not None:
+            embeddings_list.append(emb)
+            valid_mask.append(True)
+        else:
+            embeddings_list.append(np.zeros_like(representations[0]['embedding']))
+            valid_mask.append(False)
+
+        metadata.update(item.keys())
+
+    # remove embedding key from other keys
+    metadata.discard('embedding')
+    metadata = list(metadata)
+
+    embeddings = np.array(embeddings_list) # (N, D)
+    valid_mask = np.array(valid_mask) # (N,)
+
+    data = {
+        key: np.array([item.get(key, None) for item in representations])
+        for key in metadata
+    }
+
+    target_embeddings = []
+    source_regions = []
+    target_thresholds = []
+
+    for source_obj in source_objs:
+        if anti_spoofing and not source_obj.get("is_real", True):
+            raise ValueError("Spoof detected in the given image.")
+
+        source_img = source_obj["face"]
+        source_region = source_obj["facial_area"]
+
+        target_embedding_obj = representation.represent(
+            img_path=source_img,
+            model_name=model_name,
+            enforce_detection=enforce_detection,
+            detector_backend="skip",
+            align=align,
+            normalization=normalization,
+        )
+        # it is safe to access 0 index because we already fed detected face to represent function
+        target_representation = target_embedding_obj[0]["embedding"]
+
+        target_embeddings.append(target_representation)
+        source_regions.append(source_region)
+
+        target_threshold = threshold or verification.find_threshold(model_name, distance_metric)
+        target_thresholds.append(target_threshold)
+
+    target_embeddings = np.array(target_embeddings) # (M, D)
+    target_thresholds = np.array(target_thresholds) # (M,)
+    source_regions_arr = {
+        'source_x': np.array([region['x'] for region in source_regions]),
+        'source_y': np.array([region['y'] for region in source_regions]),
+        'source_w': np.array([region['w'] for region in source_regions]),
+        'source_h': np.array([region['h'] for region in source_regions]),
+    }
+
+    def find_cosine_distance_batch(
+        embeddings: np.ndarray, target_embeddings: np.ndarray
+    ) -> np.ndarray:
+        """
+        Find the cosine distances between batches of embeddings
+        Args:
+            embeddings (np.ndarray): array of shape (N, D)
+            target_embeddings (np.ndarray): array of shape (M, D)
+        Returns:
+            np.ndarray: distance matrix of shape (M, N)
+        """
+        embeddings_norm = verification.l2_normalize(embeddings, axis=1)
+        target_embeddings_norm = verification.l2_normalize(target_embeddings, axis=1)
+        cosine_similarities = np.dot(target_embeddings_norm, embeddings_norm.T)
+        cosine_distances = 1 - cosine_similarities
+        return cosine_distances
+
+    def find_euclidean_distance_batch(
+        embeddings: np.ndarray, target_embeddings: np.ndarray
+    ) -> np.ndarray:
+        """
+        Find the Euclidean distances between batches of embeddings
+        Args:
+            embeddings (np.ndarray): array of shape (N, D)
+            target_embeddings (np.ndarray): array of shape (M, D)
+        Returns:
+            np.ndarray: distance matrix of shape (M, N)
+        """
+        diff = embeddings[None, :, :] - target_embeddings[:, None, :] # (M, N, D)
+        distances = np.linalg.norm(diff, axis=2) # (M, N)
+        return distances
+
+    def find_distance_batch(
+        embeddings: np.ndarray, target_embeddings: np.ndarray, distance_metric: str,
+    ) -> np.ndarray:
+        """
+        Find pairwise distances between batches of embeddings using the specified distance metric
+        Args:
+            embeddings (np.ndarray): array of shape (N, D)
+            target_embeddings (np.ndarray): array of shape (M, D)
+            distance_metric (str): distance metric ('cosine', 'euclidean', 'euclidean_l2')
+        Returns:
+            np.ndarray: distance matrix of shape (M, N)
+        """
+        if distance_metric == "cosine":
+            distances = find_cosine_distance_batch(embeddings, target_embeddings)
+        elif distance_metric == "euclidean":
+            distances = find_euclidean_distance_batch(embeddings, target_embeddings)
+        elif distance_metric == "euclidean_l2":
+            embeddings_norm = verification.l2_normalize(embeddings, axis=1)
+            target_embeddings_norm = verification.l2_normalize(target_embeddings, axis=1)
+            distances = find_euclidean_distance_batch(embeddings_norm, target_embeddings_norm)
+        else:
+            raise ValueError("Invalid distance_metric passed - ", distance_metric)
+        return np.round(distances, 6)
+
+    distances = find_distance_batch(embeddings, target_embeddings, distance_metric) # (M, N)
+    distances[:, ~valid_mask] = np.inf
+
+    resp_obj = []
+
+    for i in range(len(target_embeddings)):
+        target_distances = distances[i] # (N,)
+        target_threshold = target_thresholds[i]
+
+        N = embeddings.shape[0]
+        result_data = dict(data)
+        result_data.update({
+            'source_x': np.full(N, source_regions_arr['source_x'][i]),
+            'source_y': np.full(N, source_regions_arr['source_y'][i]),
+            'source_w': np.full(N, source_regions_arr['source_w'][i]),
+            'source_h': np.full(N, source_regions_arr['source_h'][i]),
+            'threshold': np.full(N, target_threshold),
+            'distance': target_distances,
+        })
+
+        mask = target_distances <= target_threshold
+        filtered_data = {key: value[mask] for key, value in result_data.items()}
+
+        sorted_indices = np.argsort(filtered_data['distance'])
+        sorted_data = {key: value[sorted_indices] for key, value in filtered_data.items()}
+
+        num_results = len(sorted_data['distance'])
+        result_dicts = [
+            {key: sorted_data[key][i] for key in sorted_data}
+            for i in range(num_results)
+        ]
+        resp_obj.append(result_dicts)
+    return resp_obj

deepface/modules/verification.py

@@ -304,18 +304,21 @@ def find_euclidean_distance(
     return np.linalg.norm(source_representation - test_representation)
 
 
-def l2_normalize(x: Union[np.ndarray, list]) -> np.ndarray:
+def l2_normalize(
+    x: Union[np.ndarray, list], axis: Union[int, None] = None, epsilon: float = 1e-10
+) -> np.ndarray:
     """
     Normalize input vector with l2
     Args:
         x (np.ndarray or list): given vector
+        axis (int): axis along which to normalize
     Returns:
-        y (np.ndarray): l2 normalized vector
+        np.ndarray: l2 normalized vector
     """
     if isinstance(x, list):
         x = np.array(x)
-    norm = np.linalg.norm(x)
-    return x if norm == 0 else x / norm
+    norm = np.linalg.norm(x, axis=axis, keepdims=True)
+    return x / (norm + epsilon)
 
 
 def find_distance(
@@ -341,7 +344,7 @@ def find_distance(
         )
     else:
         raise ValueError("Invalid distance_metric passed - ", distance_metric)
-    return distance
+    return np.round(distance, 6)
 
 
 def find_threshold(model_name: str, distance_metric: str) -> float: