Two Completely Different Problems
Face detection is the task of locating faces in an image: drawing a bounding box around each face and identifying its position, regardless of whose face it is. Face recognition is the distinct task of determining whose face is in that bounding box, matching it against a known identity or finding the closest match in a database. These are separate problems with different inputs, different algorithms, and different failure modes.
You can have excellent detection and poor recognition, or vice versa. A system might locate every face in a group photograph (good detection) but fail to distinguish between two similar-looking individuals (poor recognition). Or it might recognise familiar faces with high accuracy but miss small or partially obscured faces in a scene (poor detection). The two tasks run sequentially: detection must succeed before recognition can attempt anything.
How Face Detection Works
Modern face detection uses networks such as MTCNN (Multi-Task Cascaded Convolutional Network) or RetinaFace. These networks are trained to predict bounding boxes, confidence scores, and facial landmark positions for every face in an image simultaneously. They use a cascade of progressively more complex stages, each one rapidly filtering out regions of the image that clearly do not contain faces, so the expensive final stage only needs to process regions with high prior probability of containing a face.
Facial landmarks, eye corners, nose tip, mouth corners, chin, are the key output. Ollie uses these landmarks not just for detection but for face alignment: geometrically transforming the detected face crop so that it is front-facing, centred, and standardised to the input size the recognition network expects. This alignment step is critical and often overlooked: the recognition network was trained on aligned crops, and providing misaligned input substantially degrades its performance.
Interpreting Errors
Understanding the detection-recognition pipeline helps interpret errors. If Ollie cannot find a face in your photo, that is a detection failure, most commonly caused by an extreme angle (the face must have visible landmarks), heavy occlusion (glasses covering most of the eye region, a hat covering the forehead), very low resolution, or a very small face that occupies a small fraction of the image.
If Ollie finds your face but returns surprising results, that is a recognition issue, typically related to photo conditions affecting the quality of the computed embedding. A heavily shadowed face, a wide-angle distorted selfie, or a low-quality compressed image may all produce detection success but poor recognition accuracy. The two failure modes require different diagnostic approaches.