Why traditional custom detection is slow (and costly)
Traditional computer vision projects stall because collecting and labelling thousands of real-world images takes weeks. Meanwhile, the operational question—“Can we find this specific trolley, RV or e-scooter?”—remains unanswered. The result: missed windows, bloated budgets and a sceptical board.
A breakthrough approach: Fyma’s computer vision models
Fyma’s approach is different. With about ten representative images, we can train and deploy a custom detector – often within 24–48 hours – using a rigorous synthetic data pipeline. Here’s how it works.
Step 1: Curate ~10 seed images that explain the “needle”
We start with a small, carefully chosen set that covers the essence of the object or behaviour: a luggage trolley under different angles, a branded RV, a particular waste container, a mobility scooter with a basket. Variety beats volume. If you can provide a few edge cases (partial occlusion, night lighting), even better.
For those who don’t have images – no problem: we can often capture frames from your existing camera feeds.
Step 2: Generate synthetic data—thousands of variations
Instead of waiting weeks for real-world examples, we synthesise them. Using the seed images, Fyma’s pipeline creates controlled variations: scale, rotation, backgrounds, lighting, weather, partial occlusion and motion blur. This gives the model the “muscle memory” to recognise the object in your actual environment.
Why synthetic works: it expands coverage of rare conditions (rain on glass, dusk glare) and eliminates bias towards a single viewpoint or time of day.
Step 3: Train, validate and benchmark—fast
We split synthetic and real frames into training and validation sets, add a small number of real negatives (lookalikes that are not the target), and train an object detector optimised for your camera perspectives. Validation happens both offline (precision/recall, confusion matrices) and online against live feeds to check for drift.
Governance: We track versions, metrics and ROI regions of interest (ROIs) so you know where and how the detector is used.
Step 4: Deploy to production and monitor performance
Deployment is push-button via Fyma’s platform. You can see detections and counts in context, draw ROIs and export metrics to your BI tools. We monitor precision/recall by zone and time of day, flag anomalies, and iterate quickly if conditions change.
Step 5: Iterate with active learning
As the system runs, it surfaces “hard examples.” A small set of these gets re-labelled (without PII) and folded back into the training set. This is how detectors stay robust when signage changes, layouts shift or seasonal lighting arrives.
Real-world examples
Airport operations: detect and count specific trolleys and queue states to balance lanes.
Legal & General use case: asset-specific amenity tracking to validate upgrades.
RV detection: run counts at entrances to plan parking allocation and wayfinding.
Micro-mobility: identify e-scooters vs bikes to improve storage and policy signage.
Accuracy note: Fyma has been independently tested against Avigilon (Motorola Solutions) to achieve 97% accuracy – and above. Where formal benchmarks are needed, we share methods and run controlled pilots.
Why this matters to CRE and complex estates
Speed to value: answer narrow operational questions this week, not next quarter.
Lower data burden: no multi-month labelling effort.
Flexibility: track the things that actually change your decisions, not just generic classes.
Portfolio consistency: roll the same detector across sites and compare like with like.
Privacy: detectors output counts and timings; no identity data is stored or shown.
Choosing good seeds: a quick checklist
Cover front, side and partial views.
Include scale (close and far) and lighting (day, evening).
Capture context (on the floor, by a wall, near glass).
Add negatives: objects that are similar but not the target.
Ensure image quality is representative of your cameras.
The bottom line is that with synthetic data and a tight feedback loop, custom detection becomes a two-day sprint – not a two-month project.