detect

Find named objects without training a detector

Send a photograph and the labels to find, then receive pixel boxes from google/owlv2-base-patch16-ensemble through the SIE SDK. Change the labels when the inspection job changes. The model is also available to self-host.

This example reproduces the recorded comparison behind superlinked.com/detect. It uses 2,895 seeded test images across the 100 RF100-VL datasets, with each dataset's class names as the only prompt. The domains include equipment, retail inventory, aerial imagery and scientific images. No arm is fine-tuned for this study.

Reproduce the recorded results

The code lives here. Licensed inputs, annotations, recorded outputs and usage live in the public task-evidence dataset, at the immutable revision and manifest digest pinned in fetch.py and score.py. Fetching and scoring send no model requests and require no API key.

uv sync
uv run python fetch.py
uv run python score.py

The scorer replays COCO box AP separately for each dataset, then averages the 100 scores. It also replays paired dataset-bootstrap intervals, verifies the recorded token usage, checks every displayed box against its original output, and checks the annotated-object hit counts. Altered files, duplicate images, missing rows or mismatched published figures cause a nonzero exit.

--summary-only verifies the immutable manifest, row identities, recorded score means, displayed boxes and prices, while skipping COCO and bootstrap replay. It is a quicker integrity check, rather than a replay of raw-box AP.

Arm	Mean box AP	AP50	Real-time cost per 1,000 images
SIE OWLv2 Base	11.1%	19.1%	$0.24
GPT-6 Luna	13.3%	24.4%	$0.15
GPT-5.4 mini	5.4%	13.0%	$1.33
Claude Haiku 4.5	0.9%	1.8%	$1.95
OWLv2 Large	12.2%	21.2%	Self-hosted cost depends on GPUs
OWLv2 Base, transformers	10.9%	18.8%	Self-hosted cost depends on GPUs
Grounding DINO Base	7.5%	11.5%	Self-hosted cost depends on GPUs
LLMDet Large	10.8%	16.7%	Self-hosted cost depends on GPUs

These are COCO AP percentages, not the percentage of objects found. SIE's served OWLv2 beats GPT-5.4 mini and Haiku on both AP and AP50 with positive paired 95% intervals. GPT-6 Luna scores higher and costs less. The study does not establish superiority over general vision models or fixed-taxonomy APIs. Different labels and images can change the result.

The price comparison uses each provider's recorded tokens and its real-time list rates. The displayed SIE price rounds up; competitors round down. OpenAI and Anthropic also offer asynchronous Batch rates at half their real-time prices. Both are retained in the recorded evidence.

Try a single detection request

Set SIE_API_KEY for hosted inference. To use your own SIE server, set SIE_BASE_URL to its gateway URL. Inspect the request before sending it:

uv run python run.py --show countingpills-12
uv run python run.py --case countingpills-12

Only the second command sends a paid request. It runs the same OWLv2 checkpoint, labels and 0.1 score threshold used by the recorded product check. The runner checks the catalog checkpoint before sending the image and reports the distinct execution bundle/config digest observed on the response.

Data and protocol

RF100-VL source: LibreYOLO/rf100-vl, revision 1987e22ed542539fb3d0b8a3456455c2725079a1. Every dataset's recorded MIT licence and original project attribution is in licenses.json. The export contains the sampled test annotations, selected original images, outputs from all completed main-set arms, and the sample identities. protocol.json records the decision rules, coordinate handling, thresholds and price assumptions.

The 30-image cap per dataset and seeded sampling were fixed before the main run. LLMs received the same resized images with pixel-coordinate prompts, strict schemas and no reasoning effort. Outputs were mapped back to the original image. Detector thresholds match their shipped settings. The full run, including models that outperform the served model, remains in the recorded evidence.

The paired latency check used 200 images after five discarded warmups, one request in flight, and the same Modal us-east client location for all three services. Median full-response time was 0.716 seconds for hosted OWLv2, 1.372 for GPT-6 Luna, and 1.245 for GPT-5.4 mini. The SIE/Luna median ratio was 0.522 (paired 95% interval 0.492–0.566). The registered requirement for a threefold speed claim failed; this example makes no such claim. Full raw responses and the protocol are included in the immutable evidence export.