EagleEye vs PhotonVision — Exact-frame Raspberry Pi benchmark (Dark)

Executive result

Best provisional tag match80.03%EagleEye; PhotonVision d1 81.59%; d2 75.35%

Fastest mean pipeline5.91 msEagleEye; 7.39× faster than d1, 3.12× than d2

Best mean pose error4.7 cmEagleEye; PhotonVision d1 6.0 cm; d2 6.4 cm

Frames scored22,0807 clips; complete, no timeout

EagleEye Temporal produced poses on 89.91% of frames versus PhotonVision d1’s 88.56% and d2’s 86.15%. EagleEye remained 7.39× faster than d1 and 3.12× faster than d2. EagleEye had the lowest mean 3D pose error at 4.7 cm and a 33.0 cm P99, followed by PhotonVision d1 at 6.0 cm and d2 at 6.4 cm.

Detection accuracy

The dataset marks visible tags as provisional, not official visibility truth. Accordingly, the principal number is a match-rate proxy rather than certified recall. A match requires the correct ID and corners within 5 px.

PhotonVision d1

81.59%

PhotonVision d2

75.35%

EagleEye Temporal

80.03%

System	Matched tags	Truth denominator	Proxy match rate	Mean corner error
PhotonVision d1	41,739	51,157	81.59%	0.268 px
PhotonVision d2	38,549	51,157	75.35%	0.283 px
EagleEye Temporal	40,939	51,157	80.03%	—

Processing performance

Pipeline-only wall time is compared; video decode, camera transport, network tables, UI, and preview encoding are excluded. Lower is better.

PhotonVision d1

43.67 ms

PhotonVision d2

18.42 ms

EagleEye Temporal

5.91 ms

System	Mean	Median	P95	Approx. mean throughput
PhotonVision d1	43.67 ms	45.49 ms	63.70 ms	22.9 FPS
PhotonVision d2	18.42 ms	18.46 ms	26.42 ms	54.3 FPS
EagleEye Temporal	5.91 ms	3.57 ms	21.91 ms	169.3 FPS

Pipeline inference time across the benchmark

EagleEye Temporal PhotonVision d1 PhotonVision d2 EagleEye no pose PhotonVision d1 no pose PhotonVision d2 no pose

Faint lines are individual-frame measurements; solid lines are centered 31-frame rolling medians. Scroll horizontally to inspect every frame.

Pose accuracy

EagleEye availability89.91%19,853 frames

PhotonVision d1 availability88.56%19,553 frames

PhotonVision d2 availability86.15%19,022 frames

EagleEye mean 3D error4.70 cmP95 20.0 cm

PhotonVision d1 mean error6.02 cmP95 21.3 cm

PhotonVision d2 mean error6.43 cmP95 22.5 cm

3D translation pose loss across the benchmark

EagleEye Temporal PhotonVision d1 PhotonVision d2 EagleEye no pose PhotonVision d1 no pose PhotonVision d2 no pose

Faint lines are individual-frame measurements; solid lines are centered 31-frame rolling medians. Scroll horizontally to inspect every frame.

Interactive synchronized replay

Clip Frame Frame 0

Camera view

Top-down field

Ground truth EagleEye Temporal PhotonVision d1 PhotonVision d2

Inference time overview

Pose loss overview

Method and reproducibility

Pose eligibility: Both systems permit localization from one mapped tag. EagleEye’s minimum_detections gate was changed from 2 to 1 for this rerun; PhotonVision uses production MultiTag output with its lowest-ambiguity single-tag fallback.
Hardware: Raspberry Pi Compute Module 5, ARM64, four logical CPUs, 8 GB RAM; Debian Trixie.
Dataset: seven clips, 22,080 frames, 1280×800; every system received the exact same decoded BGR frames.
EagleEye: full-dataset current-source EagleEye Temporal run. Tiny ROIs below 32 px use decimation 1; near-tied single-tag IPPE candidates use recent capture-timed pose continuity without held or smoothed output. The release temporal-projector SHA-256 is 6a7dfc755e9c44523921c0b602aa0ce39c7dd5cfc1f01b06865d51049474f231.
PhotonVision: commit 81c6aafee5b7921fc4e876044dd949dd04675815; AprilTag threads=4, refine edges enabled, decision margin=35. The report compares decimate 1 and decimate 2.
Thermals: Photon run started at 41.7°C and ended at 49.4°C; throttling flags 0x0, 0x0.
Boundary: Synthetic replay excludes physical camera transport, the application's outer threaded pipeline loop, and preview encoding/display work.

Timing includes each implementation’s AprilTag pipeline and pose work but excludes decode. PhotonVision’s first-frame startup outlier is retained. The reported throughput is simply 1000 ÷ mean pipeline milliseconds and is not camera end-to-end FPS.