Bug: CompreFace incorrectly identifies unknown faces as known subjects with abnormally high similarity scores #1310
Description
Test Scenario:
Face database: 2 persons (F000003 and F000005), each with 3 sample photos
Test image: A group photo with 5 people
Expected result: Should recognize only the 2 people in the database, the other 3 should not be recognized or have low similarity
Actual Result:
The API returned recognition results for all 5 faces, but all 5 faces were incorrectly identified as the 2 people in the database:
4 faces identified as F000005 (similarity 0.92211-0.99937)
1 face identified as F000003 (similarity 0.95638)
Problem Analysis:
This is clearly unreasonable because:
There are 5 people in the photo, but only 2 in the database - theoretically only 2 should be recognized
All 5 faces were identified as people in the database with high similarity (>0.9)
The same person (F000005) was identified as 4 different faces
Expected Behavior:
Only faces existing in the database should be recognized
Unknown faces should return low similarity or not be recognized
The same person should not be repeatedly recognized as multiple faces
Suggestion:
This might be an issue with face detection or feature extraction. Please check:
Whether the face detection algorithm accurately distinguishes different faces
If the feature extraction model is overfitting
The correctness of similarity calculation logic
测试场景：
样本库：包含 2 个人（F000003 和 F000005），每个人有 3 张样本照片
测试图片：一张 5 个人的合影
预期结果：应该识别出样本库中的 2 个人，其他 3 个人应该不被识别或相似度很低
实际结果：
API 返回了 5 个人脸的识别结果，但所有 5 个人脸都被错误地识别为样本库中的 2 个人：
4 个人脸被识别为 F000005（相似度 0.92211-0.99937）
1 个人脸被识别为 F000003（相似度 0.95638）
问题分析：
这明显是不合理的，因为：
照片中有 5 个人，但样本库只有 2 个人，理论上最多只能识别出 2 个人
所有 5 个人脸都被识别为样本库中的人，且相似度都很高（>0.9）
同一个人（F000005）被识别为 4 个不同的人脸
期望行为：
只有样本库中存在的人脸被识别
未知人脸应该返回很低的相似度或不被识别
同一个人不应该被重复识别为多个人脸
建议：
这可能是人脸检测或特征提取方面的问题，建议检查：
人脸检测算法是否准确区分了不同的人脸
特征提取模型是否过度拟合
相似度计算逻辑是否正确