Road camera anomaly watch stage 2 · open-set
road 0.98 · segmented
cone 0.96
debris 0.91
81
classes, COCO 80 plus cones
4
sizes S / M / L / X
2-stage
road, then anomalies
COCO kept
original classes retained after adding yours
Sim-to-real
synthetic data via CARLA UE 5.5 & Cosmos
What we build

Two capabilities, two public proofs

Detection when you can name what you're looking for. Segmentation when you can't, because the whole point is catching what nobody predicted. The builds below are demonstrations, not the catalog.

Custom detection Live

The capability is adding your classes to a proven detector without losing the ones it ships with. The public proof is YOLOX extended with traffic cones, published in four sizes from an edge-camera S to a server-grade X. Weld defects, missing parts, or PPE follow the exact same run.

See the full benchmark →
Road-scene segmentation Pilot

The capability is learning what normal looks like and flagging whatever breaks it. The public pilot watches motorways, where a load slides off a flatbed and gets flagged seconds after it lands. The same pattern guards any corridor, a conveyor belt, a rail track, a runway.

See the pipeline →

Where teams point these capabilities

Manufacturing QC PPE and safety zones Roads and runways Rail corridors Logistics and inventory Your cameras
Proof one · Custom detection

Cones added, COCO kept, measured

Here is the first capability with numbers attached. Four published checkpoints extend YOLOX with a traffic-cone class while keeping COCO, so you can see exactly what adding a class costs. Pick the size that matches your hardware.

Checkpoint Params COCO mAP (kept) Cone AP Where it fits Weights
yolox-pylon-s 9.0M 41.6 74.5 Edge cameras and embedded boards, real time on a Jetson Hugging Face →
yolox-pylon-m 25.3M 46.8 77.5 The accuracy/speed sweet spot for most fixed cameras Hugging Face →
yolox-pylon-l 54.2M 48.5 78.6 Server-side inference over multiple streams at 3.7 ms per frame on an A100 Hugging Face →
yolox-pylon-xl 99.1M 50.0 78.8 Maximum accuracy for offline analysis and auto-labeling, 6.0 ms per frame on an A100 Hugging Face →

COCO mAP is measured after cone adaptation and shows the retained score on the original 80 classes, within 1.2 points of the official YOLOX val2017 baseline at every size and above it at s. Cone AP is at IoU 0.50 to 0.95 on the held-out cone split. See the full benchmark →

Cones are the demo class. The same run adds your classes, whether weld defects, missing parts, PPE, or whatever your cameras need to see.

Add your classes
Proof two · Segmentation

Hazard detection without a hazard list

The second capability, running as a pilot on motorway footage. No training set can enumerate everything a road will throw at a camera, so the pipeline learns the road itself and treats anything that interrupts it as a detection. The rarer and stranger the hazard, the more clearly it stands out.

Motorway dashcam frame with the drivable road segmented in blue and a fallen box flagged in red

A frame from the pilot, captured on the M25. Stage 1 has painted the drivable corridor blue, cut cleanly around the vehicles on it. Stage 2 has flagged the box tumbling off the flatbed in red, along with the smaller fragments already scattered on the asphalt. None of these objects come from a training taxonomy. The pipeline flags them because they sit on the road and are not road.

1
Segment the corridor

SAM and SegFormer isolate the drivable road surface from the camera frame, lane by lane, in changing light and weather.

2
Flag what doesn't belong

Anything on the segmented surface that isn't road gets flagged, whether a person, a box, a cone, or a tire. The stage is open-set, so the object doesn't need to be in any training taxonomy.

3
Feed it synthetic rare events

Real debris footage is rare by definition. CARLA on Unreal Engine 5.5, SAM 3D, and Cosmos generate the long tail, labeled scenes of things that almost never happen on camera.

Built on SAM, SegFormer, and Cosmos, and currently in pilot on road-monitoring footage. The same corridor-plus-anomaly pattern transfers to conveyor belts, rail tracks, and runways.

Talk about your corridor
Why both work

One method behind both proofs

The detectors and the pipeline are not separate tricks. Both come from the house method that already runs our language and physics lines. Adapt a proven base, keep what it knows, and ship the weights with the data.

1
Start from a proven backbone

YOLOX for detection and SAM and SegFormer for segmentation. Published, benchmarked bases, never an architecture from scratch.

2
Build the labeled set

Your camera footage, labeled, and extended with synthetic scenes from CARLA UE 5.5 and Cosmos where real examples are too rare to collect.

3
Adapt without forgetting

New classes join the detector while COCO stays in the mix. The cone checkpoints keep the original 80 classes at full strength.

4
Ship weights + labels

Checkpoint and labeled dataset together, in the size your hardware wants, from a Jetson at the camera to a server GPU.

Beyond the demos

Already earning its keep on Simvera

The cone class isn't an academic demo. It's the same object family running on Simvera, our industrial perception platform, where detectors trained purely in simulation deploy to real factory and road cameras.

Trained in simulation

Photogrammetry-captured objects rendered into synthetic scenes, with no manual labeling and no waiting months for rare events to happen in front of a camera.

Deployed on real cameras

At Simvera's industrial testbed, a detector trained entirely on synthetic data localizes real traffic cones to 6.8px mean accuracy on live camera feeds. The domain gap, closed.

See it working on simvera.net

Different classes, cameras, or a private set?

The cone checkpoints prove the recipe. Your defect taxonomy, corridor, or camera network is a labeling campaign away.

Book a technical call