Case study / 07

Focal

Training a face detector from raw webcam data to real-time localization.

Focal builds face detection end to end: collecting and annotating webcam images, augmenting a small dataset, training a shared CNN backbone to jointly predict face presence and bounding-box coordinates, then running real-time inference through OpenCV.

TensorFlow
VGG16
OpenCV
Albumentations
LabelMe
Bounding Boxes

GitHub

Custom datasetDual-head modelOpenCV inference

Custom data

Collect and annotate the training set instead of wrapping a packaged face detector.

Shared backbone

One VGG16 feature extractor supports classification and localization heads.

Joint objectives

Classification and bounding-box regression are optimized in one training step.

Real-time inference

One forward pass returns face presence and normalized box coordinates.

Pipeline

Detection pipelineOwned end to end, from raw frame to rendered box

01Webcam captureOpenCV frames

02LabelMemanual boxes

03Train / test / valmanual split

04Albumentations60× expansion

05TF data pipelineimages + labels

ImageNet-pretrained VGG16shared feature backbone

07AClassification headface present?

07BBBox regression headx1 · y1 · x2 · y2

08Custom joint losslocalization + 0.5 × classification

09OpenCV inferencethreshold-gated render

Engineering decisions

Decision	Why	Tradeoff
Shared backbone with separate heads	Classification and localization can reuse the same learned image features.	Both objectives influence one shared representation.
Normalized bounding-box coordinates	Regression stays independent of image resolution and composes with bbox-aware augmentation.	Coordinates must return to pixel space at inference.
Custom training loop	The model explicitly combines localization and classification losses and controls each update.	More code and responsibility than default Keras multi-output training.

Under the hood

Data / annotation

Approximately 90 custom webcam captures
LabelMe bounding-box JSON
Manual train / test / validation split
60 augmented variants per source image

Model

ImageNet-pretrained VGG16 backbone
Sigmoid face classifier
Four-value sigmoid box regressor
Coordinates normalized to [0, 1]

Training / inference

Custom train_step and test_step
Localization + 0.5 × classification loss
OpenCV webcam input
0.5 threshold gates box rendering

Technical deep dives

Deep dive A

One forward pass. Two outputs.

The VGG16 backbone computes one shared feature representation. Independent dense heads then answer two different questions: whether a face is present and where its normalized bounds lie.

Shared features, independently shaped outputs

Deep dive B

Data becomes a detector

Bounding-box aware transforms keep labels aligned with altered images. If a crop removes the face, the pipeline writes a class-0 example and a zero box. The notebook saves a model locally; that generated artifact is not committed.

*Generated locally during the notebook run

Verified notebook behavior

10 epochs · total loss 0.35 → 0.002

Training and validation loss converge in the included run, with a transient validation spike around epoch 7. The repository does not report general detection accuracy.

Reproducibility + evaluation limits

Small, single-subject dataset from one environment.
Exactly one box output creates a one-face assumption.
No IoU or mAP evaluation against held-out ground truth.
Annotation and train/test/validation splitting are manual.
VGG16 is relatively heavy for real-time inference.
The trained model and original live video are no longer available in the repository.