Measured 2026-10-13 and 2026-10-14 · Apple M4 Max · CPU
Measured, not estimated
The numbers come from scripts/e2e.sh and retrace eval on the sample video, with SigLIP 2 B/16 at 256 px. The golden set is in the repository. We checked each hit against the crop by eye. You can run the same commands.
| What | Result | Note |
|---|
| Analysis of 79 s of video | 33 s2.4× real time | 318 frames at 4 fps, 3,121 detections, 35 tracklets |
| Search, end to end | 14 msp50 · p95 15 ms | embed the question, exact KNN, threshold, JSON. A Linux container with 2 cores: 50 to 65 ms |
| Recall at 5, seven queries | 0.95mAP 0.95 | the one miss is a white van that the detector classed as a car |
| EmbeddingGemma 2 in the process | 2 min 15 sindex · model load included | then 9 to 21 ms for each search; the binary is 57 MB |
| Bytes sent to a cloud | 0 | the model weights download one time; after that, no network is necessary |
$ retrace --embedder siglip2-b16-256 eval --golden eval/golden.vtest.json --k 5
a person in a red jacket recall@5=1.00 AP=1.00 top=[21, 30, 6, 26, 31]
a white van recall@5=0.67 AP=0.67 top=[5, 16]
two people walking together recall@5=1.00 AP=1.00 top=[7, 31, 28, 26, 23]
a bicycle recall@5=1.00 AP=1.00 top=[35]
person in a white hooded jacket recall@5=1.00 AP=1.00 top=[11, 6, 21, 31, 14]
a man in a blue jacket recall@5=1.00 AP=1.00 top=[31, 10, 24, 28, 6]
person in dark clothes recall@5=1.00 AP=0.96 top=[17, 25, 3, 18, 8]
queries=7 recall@5=0.952 mAP=0.946 p50=14.3ms p95=15.3ms embedder=siglip2-b16-256
Tracklet ids are the same for each run on one machine. A different ffmpeg or onnxruntime build can find one track more or less, which moves the ids. The end-to-end script then prints the eval without the recall check.
