Computer Vision · Deep Learning · Academic Project
Facial Keypoint Detection
Learning to locate the small points that give a face its structure—and imagining how those signals could support more human-aware systems.
01 · Purpose
Reading structure before emotion
A face carries information through very small changes: the position of the eyes, the curve of the mouth, and the relationships between individual landmarks. I built this project to understand how a deep learning model could learn those spatial relationships and predict facial keypoints from an image.
The immediate task was landmark localization. The larger idea was more human-centered: reliable keypoints can become the first layer of systems for expression analysis, attention monitoring, and emotional AI.
02 · The problem
Faces rarely appear under perfect conditions
The dataset contained 96 × 96 grayscale images and the x and y coordinates of 15 facial keypoints. The challenge was not simply finding a face. The model had to estimate precise locations while dealing with natural variation between people.
Changes in pose, expression, lighting, and image quality make the task difficult. The project documentation also identified reliability and vulnerability as important concerns.
03 · Approach
Learning spatial patterns with residual connections
I used convolutional neural networks to learn visual features from the images and residual blocks to improve gradient flow through the network. Skip connections helped the model reuse earlier activations while deeper layers learned more complex patterns.
I also applied image augmentation, including horizontal flipping, to increase variation in the training data and improve generalization. The workflow covered image preparation, visualization, augmentation, model compilation, and training in Keras with TensorFlow 2.0.
04 · What it taught me
The data pipeline shapes the model
This project made computer vision feel tangible. Plotting the target points directly on the images showed how labels, preprocessing, and model output connect. It also showed me that improving a model is not only about adding layers; augmentation, regularization, and careful evaluation matter just as much.
The most valuable lesson was learning to view the model as one part of a larger visual understanding pipeline.
05 · The next question
From landmarks to emotional AI
The original project proposed extending facial keypoint detection toward emotion analysis. A future version could use landmark movement over time, rather than a single image, to study expressions and behavioral signals.
I would also want to test broader and more diverse datasets, compare modern lightweight vision architectures, and examine fairness and reliability before using such a system in sensitive settings such as driver monitoring or affective computing.