Spaces:
Sleeping
Sleeping
| """In-memory FAISS gallery assigning persistent person IDs to face embeddings. | |
| Duplicated from ../app/identity_store.py for this standalone web demo deployable | |
| (see face_engine.py for why). Here, one instance lives per browser session | |
| (held in gr.State), not per process - nothing persists between visitors. | |
| """ | |
| from typing import List | |
| import faiss | |
| import numpy as np | |
| class IdentityStore: | |
| """Cosine-similarity gallery over L2-normalized ArcFace embeddings. | |
| Uses FAISS IndexFlatIP (inner product), which equals cosine similarity | |
| when all stored and query vectors are L2-normalized. | |
| """ | |
| def __init__(self, embedding_dim: int = 512, match_threshold: float = 0.38) -> None: | |
| self.embedding_dim = embedding_dim | |
| self.match_threshold = match_threshold | |
| self.index = faiss.IndexFlatIP(embedding_dim) | |
| self._next_person_id = 1 | |
| self.person_ids: List[int] = [] | |
| def match_or_register(self, normed_embedding: np.ndarray) -> int: | |
| """Finds the closest known identity; registers a new one if none is close enough.""" | |
| query = normed_embedding.astype(np.float32).reshape(1, -1) | |
| if self.index.ntotal > 0: | |
| similarities, indices = self.index.search(query, 1) | |
| best_similarity = float(similarities[0, 0]) | |
| best_index = int(indices[0, 0]) | |
| if best_similarity >= self.match_threshold: | |
| return self.person_ids[best_index] | |
| person_id = self._next_person_id | |
| self._next_person_id += 1 | |
| self.index.add(query) | |
| self.person_ids.append(person_id) | |
| return person_id | |
| def __len__(self) -> int: | |
| return self.index.ntotal | |