Facial Keypoints Detection

What I Explored

I explored facial landmark detection using the Kaggle Facial Keypoints Detection dataset. Unlike ordinary image classification, the goal is not to assign a class to an image but to predict spatial coordinates corresponding to important facial landmarks.

This introduced me to computer vision as a regression problem, where a CNN learns visual features and maps them to continuous coordinate values.

What Confused Me

One of the interesting challenges was understanding how a network can move from extracting spatial image features to predicting precise numerical coordinates.

Handling incomplete landmark annotations and choosing appropriate preprocessing and loss functions also made the problem different from standard classification.

What Clicked

I realized that convolutional networks are not inherently limited to classification. The feature representation produced by convolution layers can support many downstream objectives.

By changing the output representation and loss function, essentially the same feature-extraction principles can be used for classification, localization, regression, and detection.

My Mental Model Now

I think of facial keypoint detection as:

Image → Visual Features → Spatial Representation → Coordinate Regression

The CNN progressively identifies patterns such as edges, facial structures, eyes, noses, and mouths. The final layers transform these learned features into landmark coordinates.

Implementation / Example

I built a notebook-based training pipeline for the Kaggle Facial Keypoints Detection problem and saved the trained PyTorch model for subsequent inference.

Repository: Facial Keypoints Detection

What I Want to Explore Next

I want to explore heatmap-based landmark prediction, modern pose-estimation architectures, stronger augmentation strategies, and how landmark localization differs from bounding-box-based object detection.