Street Scene Vision Toolkit

How to use the app

1 · Choose an image

Upload, paste, use a webcam, or select the UBC campus example.

2 · Segment, then detect

Start with SegFormer semantic segmentation, then optionally run YOLO object detection on the same image.

3 · Explore and download

Compare boxes, overlays, masks, counts, pixel shares, coordinates, and reusable CSV outputs.

What the results mean

Object detection finds separate COCO objects, draws bounding boxes, and reports a confidence score for each detection.
Detection indicators summarize visible people, active-mobility objects, and transport objects. They are transparent image counts, not traffic-flow estimates.
Segmentation overlay blends the Cityscapes prediction with the original photograph. White lines mark class boundaries.
Color mask and area share show pixel-level scene composition. Area share describes visual coverage, not physical land area.
Downloadable data include bounding-box coordinates, class-ID pixels, overlays, masks, and CSV summaries.

Scene layer	Segmentation classes
Travel surfaces	road, sidewalk
Built environment	building, wall, fence, pole, traffic light, traffic sign
Nature and sky	vegetation, terrain, sky
People	person, rider
Transport	car, truck, bus, train, motorcycle, bicycle

Classroom and research ideas

Compare detected people, bicycles, and motor vehicles across several street images.
Compare what bounding boxes reveal with what pixel-level segmentation reveals.
Discuss missed objects, false positives, confidence thresholds, and segmentation boundary errors.
Export both CSV files and build object-count and class-coverage charts.
Compare the same location across seasons, weather conditions, or camera viewpoints.

Important: predictions are model estimates, not ground truth. COCO detection is limited to its trained object vocabulary, while Cityscapes segmentation is specialized for road-driving imagery. Do not use either output for safety-critical decisions, surveillance, or identifying individuals.