Wireframe scene of the camera, the field-of-view frustum, the focus planes, the reliable detection range and every person with its live pixel reading, among the scenery of the chosen scene. Figures are added and placed with the Subjects fields or dragged with the mouse; each reading is also listed in the results.
Pixels on Target
Ideal-conditions estimate from pinhole geometry alone. Lens distortion, focus blur, image processing, compression, lighting, motion, occlusion, pose and the model’s training data all shorten the real-world range.
Sensor
Sony IMX290
Two photosites: the floor below which a sampled grid cannot tell blur from sharp (Nyquist).
1920 × 1080 px · 2.1 MP · 6.39 mm diagonal · Rolling shutter · Color / NIR
Lens
4 mm · f/2.8
Focus
5 m
Detection
YOLOv5 / v8 / 11 (640 px)
Threshold basis
Heads at strides 8/16/32 on a 640 px letterboxed input. COCO scores objects under 32 × 32 px as "small", where every YOLO generation reports far lower AP; pedestrian benchmarks call ≥ 50 px tall the "reasonable" range.
Source: Ultralytics documentation (imgsz 640, letterbox resize); COCO detection evaluation (small objects < 32 × 32 px); Dollár et al., "Pedestrian Detection: An Evaluation of the State of the Art", PAMI 2012 (reasonable ≥ 50 px, near ≥ 80 px, medium 30–80 px)
Subjects
1 person
Distance along the optical axis; offset to the camera’s right. Figures can also be dragged in the scene.
Scene
1.5 m
A scene surrounds the camera with scale scenery and its own figures, which can be dragged. The lens height is measured above its ground. Both are shared with the other tools.
Detection
- Max reliable distance
- 15.63 m
- Detections begin
- 24.43 m
- On sensor at focus
- 469 px
- Delivered to model
- 156 px
- Requirement (model input)
- ≥ 50 px
- Pixel density at focus
- 276 px/m
Subjects in scene
- Person 15.0 m469 → 156 px
Field of view
- Horizontal
- 69.7°
- Vertical
- 42.8°
- Diagonal
- 77.2°
- Per pixel (IFOV)
- 0.725 mrad
- Field at focus
- 6.96 × 3.92 m
- Resolution at focus
- 3.62 mm/px
- Magnification
- 0.000801×
Depth of field
- Near limit
- 823.6 mm
- Far limit
- ∞
- Total
- ∞
- Hyperfocal
- 989.2 mm
- Diffraction limit
- f/4.3
- Airy disc
- 3.76 µm
How it works
Pixels on target
A target of height h at distance d spans n = h·f / (d·p) pixels, for focal length f and pixel pitch p (the pinhole, ground-sample-distance model). Figures in the scene are projected through the camera at its mounting height, so one off to the side or well below the camera reads a little differently from one on the axis.
Frame to model
When the frame is resized into the model's input, the count scales with it: letterbox scales by the longer side, stretch scales each axis on its own.
Thresholds and range
Each model has a marginal height where detections begin and a reliable height from its published benchmarks. The maximum reliable distance is where the target shrinks to the reliable height, d = h·f / (n·p).
Found an error? feedback@frustum.tools