Lab 01
Live Vision
See what the AI sees
Play a video and watch an AI find people and everyday objects, right on your device.
Overview
A camera produces more footage than a person can continuously review.
How this could help a business
Visual detection can be one component of monitoring, stock checks or inspection software. A business-specific system needs relevant training data and a plan for mistakes and missed objects.
What I built
- A video application supporting a sample, local files and optional camera access.
- A browser worker connecting object detection, face detection, frame-to-frame tracking and a video overlay.
Live demo
Start with the sample video. You can use your camera or a file afterwards.
What I built
Implementation, decisions and evidence.
What Sebastian built
The application around the technology.
- A video application supporting a sample, local files and optional camera access.
- A browser worker connecting object detection, face detection, frame-to-frame tracking and a video overlay.
- Device measurements, model verification and an inspector for investigating performance.
One engineering decision
Process video locally and keep rendering separate from heavy detection work so the page can remain responsive.
What to check yourself
Play the sample to see real detections. Inspect the measured frame timings and the committed accuracy benchmark in What I built.
Tools and model credits. D-FINE and MediaPipe supply pretrained detectors. This application integrates them; it does not claim to have trained them from scratch.
These are portfolio demonstrations. Explore professional experience separately, or discuss how the approach could fit your team.
Learn
Understand what the application is doing.
Look for boxes and names around objects in the video. A missed object or a wrong label is a useful way to discover the limits of the detector.
Pause and predict
If you ask the detector to be more certain, will it usually draw more boxes?
A learning question, not a score. Try this in the working demo.
Try changing something
Now make it your experiment.
Pause on a busy frame. Change the confidence setting and observe how the balance between shown and hidden boxes changes.
Objects can be missed or mislabelled, and the tracker never identifies anyone. Inference on a laptop GPU is still slower than my per-frame target; the benchmark in What I built has the measured numbers.
A few words, explained
- Confidence
- How certain the model's detection score appears. A high score is not a guarantee.
Data handling & current limitations
Video stays on your device. The first run downloads the detector; speed depends on your device.
Demo only. Each lab below explains which inputs stay on your device and which data is sent for server-backed features. Model files are downloaded and may remain in your browser cache. Hosting and security services also receive the network requests needed to serve the site.
Camera frames and video files are processed in this browser and are not uploaded. Downloaded model files may remain in your browser cache.
Objects can be missed or mislabelled, and the tracker never identifies anyone. Inference on a laptop GPU is still slower than my per-frame target; the benchmark in What I built has the measured numbers.