Our viva question bank has a short list of questions that are really aimed at one kind of project. One of them is aimed at face recognition: “What does it do with a face it has never seen?” Beside it the bank notes what the examiner is testing: rejection versus nearest-match guessing. It is the whole project in one question, which is why I start with it.
The student’s problem, in plain words
You want to build a face recognition attendance system project that runs on your own laptop, marks students present from a webcam, and survives a viva. The first part is easy to find. Search results give you tutorials that detect a face and draw a box, and some that print a name above it. What they rarely cover is what the name means when the face is not in your dataset.
Picture a student whose demo works perfectly with five classmates enrolled. This is a made-up student, but the setup is the usual one. The examiner walks in front of the camera, someone who was never enrolled, and the screen says “Rahul, present”. Nothing crashed. The system just did what the naive version always does.
Chapter one: how the matching really works
The project page describes the pipeline. At enrolment, a few photos of a student go through the face-recognition library, which is built on dlib. It finds the face and computes a 128-number encoding, and the average of those encodings becomes the student’s stored reference. During a class, OpenCV reads frames, each detected face is encoded, and the encoding is compared with every enrolled encoding by Euclidean distance.
An encoding is a vector produced by a network trained so that photos of the same person land close together and photos of different people land far apart. That sentence is the answer to “what is a face encoding?”, and it is worth being able to say it without notes.
Chapter two: nearest match always finds someone
Here is the problem hidden inside the simple version. If your code says “find the enrolled student with the smallest distance”, then there is always a smallest distance. A stranger gets matched to whoever looks least different from them. The code does not malfunction. It answers a question nobody should have asked.
The fix on the project page is a threshold. A face counts as identified only if its nearest distance is under 0.5, which is stricter than the library’s 0.6 default. Otherwise it is labelled unknown. Notice what this really is: a trade. A lower threshold means fewer false matches and more missed students. A higher one means the opposite. The project is meant to show you that trade-off with numbers rather than hide it, and the project page suggests the public Labeled Faces in the Wild dataset for testing the threshold.
That is the turning point of this project. It stops being “recognise faces” and becomes “decide when not to answer”. Every good machine learning viva eventually lands there. Our question bank says the same thing for ML projects generally: show me a case where your model is wrong.
Chapter three: speed on an ordinary laptop
The second thing that breaks on a student laptop is the frame rate. Encoding every frame at full resolution is too slow. The project downsizes frames and processes only every third one, which keeps the feed smooth without missing anyone who stays in view for more than a moment. It is a small decision with a good viva answer: you can say what you gave up (the faces that flash by in under three frames) and why that is acceptable in a classroom.
Chapter four: once per session, and the honest limit
Attendance needs one row per student per session. The project uses a unique constraint on student and session in the attendance table, so later detections in the same session are ignored. This is the same pattern as preventing a double booking in a hospital system. Let the database refuse the duplicate instead of hoping your code checks first. If you want to see that idea in a different project, read the hospital management system project story.
Then the limit you should state before the panel does. A printed or on-screen photo can fool the base version. The project page is honest about this: liveness checks such as blink detection are discussed as an extension, and the base version does not include one. Saying that yourself, calmly, does more for your viva than pretending the system is spoof-proof.
What I would do again
- Store encodings, not photos, and delete the photos after enrolment.
- Set a distance threshold and write down why you chose it, ideally tested on a public dataset.
- Show the unknown-face case in your demo before anyone asks for it.
- Downscale frames and skip frames, and say what you gave up.
- Enforce one attendance row per student per session with a unique constraint.
- Put the photo-spoofing limit in your report and name liveness detection as future work.
- Export attendance to CSV, because that is what a department actually asks for.
Viva answers to rehearse
Why encodings and not a CNN classifier per student: enrolling is just storing a vector. How do you decide two faces match: Euclidean distance below 0.5. Can someone fool it with a photo: yes in the base version, and liveness detection is the standard fix. Show me it failing in poor lighting: have a screenshot or a short recording ready, because the bank’s own note says the point is whether real-world limits were tested. The question bank has the rest, and the guide on AI projects for final year students compares this project with others in the same family.
The face recognition attendance project class, with its six live classes and what each tier includes, is listed on the face recognition attendance page. If you are not sure whether a vision project suits your degree, the find my project quiz takes a couple of minutes.