# 🎯 Face Landmarks Detection (OpenCV DNN + Facemark)

August 20, 2025

This content originally appeared on DEV Community and was authored by Ertugrul

“Faces don’t lie — but landmarks sometimes do.”

Project Idea

Hey there! In this post, I’ll share my journey of building a Face Landmark Detection pipeline using OpenCV DNN and Facemark LBF. The system takes a raw video as input, detects faces, extracts 68 facial landmarks, smooths them across frames, and finally outputs:

an annotated video with landmarks and bounding boxes
an optional CSV file with landmark coordinates for every frame

The idea was simple:

“Take a face → get the points.”

But to make it robust, I had to mix deep learning detection with classical landmarking and add a touch of signal processing.

Code Structure

The project is split into modular components:

detector.py → Loads and runs the DNN-based face detector (SSD ResNet)
landmarks.py → Drawing utilities for the 68-point facial structure
helpers.py → Video I/O, CSV logging, smoothing, and per-frame pipeline
main.py → Entry point to run the full pipeline

Step 1 — Face Detection

I used OpenCV’s Deep Neural Network (DNN) SSD ResNet model. The detector takes each frame, converts it into a blob, and feeds it into the Caffe network:

blob = cv.dnn.blobFromImage(cv.resize(frame, (300, 300)), 1.0, (300, 300),
                            (104.0, 177.0, 123.0), False, False)
self.net.setInput(blob)
detections = self.net.forward()

This gives us bounding boxes with confidence scores. I kept only the ones above a threshold (conf_thr=0.6).

Step 2 — Landmark Extraction

With face boxes ready, I used Facemark LBF to extract the 68 landmark points:

facemark = cv.face.createFacemarkLBF()
facemark.loadModel(LBF_MODEL)
ok, landmarks = facemark.fit(frame, np.array(boxes))

This returns arrays shaped (68, 2) → coordinates for jawline, eyebrows, eyes, nose, and lips.

Step 3 — Landmark Smoothing

Raw landmarks jitter a lot between frames. To stabilize them, I applied an Exponential Moving Average (EMA):

if prev_pts is None:
    smooth = pts.copy()
else:
    smooth = alpha * prev_pts + (1.0 - alpha) * pts

This keeps the motion natural but removes frame-by-frame noise.

Step 4 — Drawing the Mesh

I grouped the 68 points into face regions and connected them with polylines:

Jawline
Eyebrows
Nose bridge & base
Eyes
Inner & outer lips

for (x, y) in pts:
    cv.circle(frame, (int(x), int(y)), 1, (0, 255, 0), -1)

The result? A clear, real-time facial mesh overlay.

Outputs

Annotated Video:
Watch on YouTube
CSV Example:

  frame_idx,x0,x1,...,y66,y67
  0,123,130,...,200,205
  1,124,129,...,199,206

This makes the system useful both for visualization and downstream ML tasks.

Lessons Learned

DNN face detection is robust, but combining it with traditional landmarking is still effective.
Smoothing is mandatory — raw landmarks are too noisy for real use.
CSV logging adds value for research/analytics beyond just visualization.

GitHub Repository

You can find the full code here:

GitHub: Face Landmarks Detection

“A single face in a frame is simple — but tracking it smoothly across time is where the real challenge begins.”

This content originally appeared on DEV Community and was authored by Ertugrul

dnn python