Measured report · 02

How much accuracy does lower-resolution object detection trade away?

We measured six local YOLO11 configurations on all 5,000 COCO validation images, then checked 640-pixel deployment on Cloud Run L4 GPU, Core ML and C4 CPU.

5,000

validation images

local configurations

46.37%

selected mAP50–95

13.55 ms

selected p95 latency

The frozen rule selected YOLO11S at 640 pixels: it was the lowest-p95 configuration within 2.0 absolute mAP50–95 points of the YOLO11S 640 reference.

Its measured 13.55 ms p95 implies 36.9 aggregate frames per second after reserving 50% latency headroom. That is a capacity estimate, not a production guarantee.

Measured frontier

Resolution bought accuracy, but not proportional latency

Every accuracy value uses the full COCO 2017 validation split. Latency is end-to-end batch-one prediction after 20 warm-ups.

Scatter plot of COCO mAP50–95 versus p95 latency for YOLO11n and YOLO11s at 320, 480 and 640 pixels — The upper-left region is preferable. A higher input resolution improved accuracy for both checkpoints; measured Metal p95 did not rise monotonically at every step.

This plot checks deployability across runtimes; it is not a controlled hardware ranking. The Cloud Run L4 measured YOLO11n at 13.05 ms p95 and YOLO11s at 13.22 ms p95.

All local results

The six configurations behind the decision

Model	Input	mAP50–95	Person AP	p50	p95	p99	Safe FPS
YOLO11N	320px	28.82%	39.24%	10.34 ms	12.73 ms	17.73 ms	39.3
YOLO11N	480px	35.71%	47.29%	9.83 ms	13.71 ms	16.95 ms	36.5
YOLO11N	640px	38.88%	51.76%	7.77 ms	10.32 ms	21.64 ms	48.5
YOLO11S	320px	37.52%	46.57%	8.00 ms	10.09 ms	19.62 ms	49.6
YOLO11S	480px	43.40%	54.27%	8.87 ms	10.20 ms	15.20 ms	49.0
YOLO11S · selected	640px	46.37%	57.94%	11.32 ms	13.55 ms	23.40 ms	36.9

Business sizing

Translate latency into an inspection-rate hypothesis

At the selected p95, the 50%-headroom model supports the following evenly scheduled camera scenarios before decoding and application I/O.

cameras at 2 fps each

cameras at 5 fps each

cameras at 10 fps each

Frozen protocol

Inputs, labels and execution

Dataset: COCO 2017 validation split
Evaluator: pran-vision-deployment/1.0.0
Executed: 2026-08-24
Local hardware: Apple M4 Pro, 16-core GPU, PyTorch MPS
Selection: Lowest p95 local latency within 2.0 absolute mAP50-95 points of YOLO11s at 640 px.
Cloud execution: One NVIDIA L4 on Cloud Run in europe-west4 ran both 640-pixel PyTorch/CUDA configurations. The original us-central1 one-L4 quota preference remained pending; C4 CPU/ONNX stays separately labelled.

COCO dataset source

Download evidence

Results, method and reproduction commands

No email gate. Model hashes, environment versions and runtime-specific raw timings are included.

Configuration results (CSV)Frozen result summary (JSON)Methodology and limitations Reproduction commands Local raw timings (JSON)Cloud C4 raw timings (JSON)Cloud Run L4 raw timings (JSON)

Sizing an inspection system? Start with the site constraints and the evidence needed to clear a production gate.

Plan a vision feasibility sprint