01 / CAPTURE → UNDERSTAND → DOCUMENT
Your room.
A new perspective.
Point your camera. See what it recognises.
Build a visual record of your space.
Start with what’s around you.
Try a chair, a plant, a laptop or a bottle.
Good light. Slow movements. Real detections.
02 / PUT YOUR VIEWS IN CONTEXT
A simple capture map.
Select a saved view, then tap where you took it on the sketch. Pins mark camera positions—not detected object locations. The rectangle is a blank sketch, not a measured room.
Save a view first. Keyboard: arrow keys move the placement cursor; Enter places the selected view.
Saved views 0
Up to 24 views. Kept in this tab only—export before closing.
Your evidence trail starts here.
What works, what doesn’t, and the open-source pieces ＋
Real detection. Local processing.
TensorFlow.js + COCO-SSD analyse your camera frames on your phone. The model downloads once per browser cache (about 19 MB). Photos and clips are not sent to Northframe. No microphone is used.
A local video is sampled at up to 8 moments. Each saved view records its clip timestamp. Repeated detections across views may be the same object; this is not an inventory count.
A capture map, not a 3D scan.
You place camera positions yourself. There is no automatic camera tracking, depth measurement, wall reconstruction or accuracy guarantee. Predictions need review.
For actual video-to-3D mapping, DROID-SLAM is a research starting point requiring a separate GPU setup. Apple RoomPlan is a native alternative on LiDAR-equipped devices; the standard iPhone 14 does not have LiDAR.