Why YOLO-P2?
The P2 prediction head works at a higher spatial resolution than conventional deeper heads. It preserves shallow feature detail that is useful when pedestrians, bicycles and vehicles occupy very few pixels in wide-area imagery.
Live execution
Awaiting input
0%
Detection product
Visual result
P2
No result yetChoose an image or video. The completed product will remain fully contained inside this panel.Process telemetry
What the system is doing
Elapsed0 ms
Route—
Units—
SYSTEM READYSelect an image for first-stage validation.
No execution events.
METHOD & METRIC GUIDEUnderstand every policy, stage and reported attribute ＋
Why adaptive SAHI?
Large scenes are not equally difficult. The system measures scene complexity, then chooses full-frame inference or overlapping 640/768-pixel tiles. It spends additional computation only when the policy judges that local detail requires it.
OpenVINO INT8 versus PyTorch FP32
OpenVINO INT8 is the operational edge path: quantized weights, batch 1 and latency-oriented CPU execution. PyTorch FP32 is retained as a reference/fallback path. Backend agreement is not accuracy validation; use labelled evaluation data for that claim.
Why report-only video?
Detection and evidence generation are the useful analytical work. Re-encoding an annotated MP4 adds CPU time, disk writes and storage. Report-only mode keeps JSON, CSV and HTML evidence while avoiding that optional cost.
The 15-stage execution trail
Checks file type, upload size and whether the media can be decoded before expensive model work starts.
Loads only the requested OpenVINO INT8 or PyTorch FP32 detector. Switching backend releases the previous detector where possible.
Confirms the detector, ten-class VisDrone label map, CPU device, batch-one policy and model runtime are available.
For video, selects frames using target inference FPS and maximum stride. A strong scene change can force an unscheduled inference.
Measures texture, blur/focus, entropy, brightness and contrast on a compact grayscale proxy. These are routing signals, not detection accuracy.
Combines quality signals into a complexity score, then chooses full-frame, 768-tile or 640-tile processing using configured thresholds.
Creates an edge-complete overlapping tile grid. Overlap protects objects cut by tile boundaries but increases processed pixels.
Runs the selected detector once per chosen full frame or tile at batch one. This is normally the dominant compute cost.
Adds each tile's x/y origin to local boxes, restoring every prediction to the original full-frame coordinate system.
Applies class-wise non-maximum suppression across all slices to remove duplicate boxes that overlap in neighbouring tiles.
Calculates class, confidence, size, density, spatial, workload and NMS statistics; video additionally reports sampling and track estimates.
Draws labels and boxes for images or optional video. Report-only video can omit expensive frame drawing and MP4 encoding.
Creates a standalone HTML evidence report that documents configuration, outputs, timing and interpretation boundaries.
Saves the annotated image when applicable, evidence JSON, video detection CSV and report. Source uploads are removed by edge policy.
Closes the reproducible execution trail and exposes only the output files produced by the selected policy.
Reported attribute dictionary
Complexity scoreWeighted routing score from texture, blur and entropy. Higher values trigger finer slicing according to configuration.
Texture indexNormalized high-frequency content estimated from Laplacian response. Detail, edges and noise can all increase it.
Focus / blurLaplacian variance estimates sharpness; blur index is its normalized inverse. It is a heuristic, not an optical calibration.
EntropyIntensity-distribution information in bits. Higher entropy usually means a more varied visual scene.
Brightness / contrastMean grayscale intensity and standard deviation. They expose difficult low-light, washed-out or low-contrast inputs.
TilesNumber of full-frame or overlapping crops actually inferred. More tiles usually improve local scale but increase latency.
Processed megapixelsTotal pixels presented across all tiles, including overlap—not simply the original image resolution.
Workload amplificationProcessed tile pixels divided by original frame pixels. A value of 2.0 means twice the original pixel workload.
Candidates before NMSAll confidence-filtered boxes collected before cross-slice duplicate removal.
Suppressed by NMSBoxes removed as same-class overlap duplicates. The suppression ratio helps audit overlap and fusion behaviour.
Fresh detectionsBox instances from frames that were actually inferred. Propagated display boxes are excluded from this count.
Estimated unique objectsClass-aware IoU track count. It reduces repeated frame detections but is not a ground-truth identity count.
Small-object percentageShare of boxes with area below 32² pixels using the COCO pixel convention on the processed output coordinates.
Detection densityDetections per image megapixel, or per megapixel per inferred frame for video. Useful for comparing scene occupancy.
Confidence statisticsModel score distribution after thresholding. Confidence is not calibrated probability and should not be read as accuracy.
Box-union coveragePercentage of image pixels covered by the union of all detection boxes; overlapping box pixels count once.
Spatial quadrantsCounts assigned by box centre to top-left, top-right, bottom-left and bottom-right regions.
Inference coveragePercentage of video frames receiving fresh inference. Lower coverage reduces compute but may miss brief events.
Inference reductionTheoretical percentage of frame-level detector calls avoided by sampling compared with inference on every frame.
Scene overrideAdditional inference caused by a frame-to-frame change exceeding the configured threshold despite normal stride skipping.
P2 inference timeDetector execution time accumulated for the selected route. End-to-end time also includes upload, slicing, fusion, drawing and exports.
Validity boundary: these attributes validate execution behaviour and resource strategy. Precision, recall, mAP and missed-object claims require labelled ground truth evaluated under a fixed protocol. Estimated tracks are not verified identities, and propagated video boxes are presentation continuity rather than fresh detections.