# Detect page evidence The page shows Grounding DINO boxes on eight licensed photographs of store shelves and warehouse floors. A ninth recorded photograph, a fulfillment center tour group, appears only in this record. Every total on the page counts all nine photographs. ## Model run - Page model: [`IDEA-Research/grounding-dino-base`](https://huggingface.co/IDEA-Research/grounding-dino-base), the registry quality lane and the playground's example model - Hugging Face revision in the SIE model config: `12bdfa3120f3e7ec7b434d90674b3396eccf88eb` - Also recorded: [`google/owlv2-base-patch16-ensemble`](https://huggingface.co/google/owlv2-base-patch16-ensemble), the registry fast lane, revision `cfd3195ba4ea9592eec887ded089f4c08eff231d` in the SIE model config - `x-sie-model-revision` response header on every call: `10333b84de80b402376b626eb25366fb081d3faeb893eb4b01cf32e8c27e4aff` - Server version header: `0.7.3` - Endpoint: `https://api.superlinked.com` - Run started: `2026-09-16T03:32:37Z`, 18 calls, all HTTP 200 - Paper: [Grounding DINO](https://arxiv.org/abs/2303.05499) Each call is the request the page's playground snippet sends: one JPEG image item and the comma-separated labels, with no threshold options. Grounding DINO therefore runs at its SIE defaults (`box_threshold` and `text_threshold` 0.25) and OWLv2 at its SIE default `score_threshold` of 0.1. The runner, inputs, stored requests, raw responses, run manifest and the exact JPEG bytes sent live in `apps/site/tests/fixtures/reference/detect/`. Stored requests replace the base64 image with its SHA-256 and byte length. Re-run from the repository root with `python3 apps/site/tests/fixtures/reference/detect/run.py`. ## Boxes and display SIE returns each object's `bbox` as `[x, y, width, height]` in pixels of the image that was sent. A card showing the whole photograph converts those four numbers to percentages of the sent image. A card that trims its view first subtracts the view's origin, then divides by the view's width and height. The served WebP files are re-encoded copies of the sent JPEGs with identical pixel dimensions, so the recorded coordinates apply to them unchanged. Labels render exactly as the model returned them, which is why the soft drink photo shows `7 up bottle` for the requested `7 Up bottle`. Scores render with two decimals. The container loading dock card trims empty sky and bare yard from its display. It shows the region `x 0, y 440, width 1,050, height 890` of the sent image, and every returned box sits inside that region. The other photos display in full. The playground shows the ship's hangar deck photo in full with the exact response its snippet prints; its one missed pallet jack is not outlined there. ## Counting method Labels and hand counts were fixed in `inputs.json` before any model call, in the first commit that adds that file. A count includes every instance a person can see in the sent image, partly hidden ones included unless the case note says otherwise. After the run, a person reviewed every returned box on a pixel grid: - **found**: the first box on a counted object of that label; - **duplicate**: another box on an object that already has one; - **wrong object**: a box that does not sit on an object of its label. A box only counts for the label it carries, so a `person` box over a shopper who hides a price sign never counts as that sign. A box that spans the whole photograph locates nothing and counts as a wrong object, not as every instance it happens to contain. Counted objects without a box are outlined by hand with dashed lines on the hero and the proof cards. Those outlines are approximate display marks, not model output. The playground shows its photo with the response alone, so the one pallet jack it missed carries no outline. The tour photo is not displayed at all, so its misses are counted without outlines; two of its twelve vests are mostly hidden behind other people, and the pre-run count keeps them. Every one of the 36 counted objects with no box was then re-checked mechanically, by measuring the overlap between each recorded miss, drawn or not, and every returned box on that photograph. Fourteen looked covered. Twelve are the soft drink shelf, where the frame-sized `price tag` box contains all eleven tags and the missed bottle without locating any of them. The other two are label mismatches: a `person` box over the shopper who hides a yellow price sign, and a `person` box over the soldier pulling an unboxed pallet jack. None of the 36 changed class, and no returned box was credited to the wrong instance. ## Results | Photo | Labels sent | Counted | Found | Boxes | Duplicates | Wrong object | |---|---|---:|---:|---:|---:|---:| | Sauce aisle (hero) | sale sign | 14 | 10 | 10 | 0 | 0 | | Container loading dock | forklift, shipping container, person | 4 | 4 | 6 | 2 | 0 | | Food box packing floor | safety vest | 14 | 9 | 9 | 0 | 0 | | Produce department | person, shopping cart, yellow price sign | 17 | 11 | 12 | 1 | 0 | | Soft drink shelf | price tag, 7 Up bottle | 14 | 2 | 3 | 0 | 1 | | San Juan distribution center | forklift, pallet jack | 2 | 1 | 2 | 0 | 1 | | Javits Center staging hall | pallet jack, person | 6 | 4 | 4 | 0 | 0 | | Ship's hangar deck | pallet jack | 5 | 4 | 4 | 0 | 0 | | Fulfillment center tour (not displayed) | safety vest | 12 | 7 | 7 | 0 | 0 | | **All nine** | | **88** | **52** | **57** | **3** | **2** | 55 of the 57 returned boxes sit on an object with the label they name. The two wrong boxes are a `price tag` box covering the whole soft drink photo and a `pallet jack` box on the San Juan forklift. Small shelf-edge price tags came back 0 of 11. ## OWLv2 at its default threshold The same requests on OWLv2 returned these box counts. The page does not display them. | Photo | OWLv2 boxes | Hand count | |---|---:|---:| | Sauce aisle | 47 | 14 | | Container loading dock | 19 | 4 | | Food box packing floor | 35 | 14 | | Produce department | 378 | 17 | | Soft drink shelf | 380 | 14 | | San Juan distribution center | 11 | 2 | | Javits Center staging hall | 20 | 6 | | Ship's hangar deck | 20 | 5 | | Fulfillment center tour | 16 | 12 | At a 0.1 score threshold OWLv2 boxes nearly every shelf region, so its default output cannot be counted without a higher threshold. ## Candidates dropped before any run Three candidates were dropped while writing `inputs.json` because their counts were ambiguous: Nick Saltmarsh's Sainsbury's distribution centre photograph, Benoît Prieur's Leclerc Drive warehouse photograph and a U.S. Navy Exchange shelf photograph. None of them was sent to a model. ## Image records Each input is the Wikimedia Commons original, EXIF-rotated, optionally cropped, resized with Lanczos to 1,400 px on the longest side and saved as progressive JPEG quality 84. | Local file | Source | Creator | License | Original SHA-256 | Crop (`x1, y1, x2, y2`) | Sent image SHA-256 | |---|---|---|---|---|---|---| | `sauce-shelf.webp` | [Commons](https://commons.wikimedia.org/wiki/File:King_Soopers_Grocery_Store,_Colorado_Springs_(50842301541).jpg) | David (Flickr) | CC BY 2.0 | `f804fc29c07452e0f701be2bda97eb09fe1977bbae7874b57b9004594cae42da` | none | `84aba6b9eb806f54cd32e4136692dde7be66188b0d60a2364ec6085f19710e82` | | `container-dock.webp` | [Commons](https://commons.wikimedia.org/wiki/File:Container_loading_with_forklift_at_warehouse_in_Thailand.jpg) | Goterrestrial | CC BY 4.0 | `3f7acb0cdd2a47a56afab1058cd2049aaf8d7f67e03e05326174bc1b0fefed1f` | none | `ba81761da8d40d3d5cdf7d91863b361b476292d692e43716457abfb20afdc1b1` | | `food-box-floor.webp` | [Commons](https://commons.wikimedia.org/wiki/File:Food_box_packing_facility_during_covid.jpg) | U.S. Department of Agriculture | Public domain | `b9620d3fc7335f170f38908c8e734875add747c45a88390606e206dc52de5d39` | `3000, 1640, 5464, 3640` | `aef7930e222d5b8a6b6c8babc071a8fec2805aba14add8648eb831ab8f440ba7` | | `produce-department.webp` | [Commons](https://commons.wikimedia.org/wiki/File:Supermarket_(14961829054).jpg) | Dean Hochman | CC BY 2.0 | `b4174805a8cbc7ece1cb3b1397fb242a5424321dd037b5e5674a59e9f627cf1e` | none | `5b707509f130d326572fcb4dc64b7e82ba0943effcb6859cb5154097442b2bc8` | | `soft-drink-shelf.webp` | [Commons](https://commons.wikimedia.org/wiki/File:Soft_drink_shelf.JPG) | SMC at English Wikipedia | Public domain | `97013217c9c0441dc3aef2ab8ed1087a559d865c547d174185290d02f4d14549` | none | `dbdf9c8e3a4a50b9dfb357c6c5b88b0ee3700f4cc1a1b3aa600320b0277f0aa9` | | `fema-caribbean-forklift.webp` | [Commons](https://commons.wikimedia.org/wiki/File:Emergency_Planning_and_Security_-_San_Juan,_Puerto_Rico,_May_29,_2012_--_FEMA%27s_Logistics_Management_Division_open_the_agency%27s_new,_strategically-located_Caribbean_Distribution_Cen_-_DPLA_-_d2c4633e1ec7478e8b435ea7ddbce4ce.jpg) | FEMA, U.S. Department of Homeland Security | Public domain | `8a83989fb034cb4300375f6e4ea64b2b3927863b7c8f3581409bd8da112afbe7` | none | `0587dc67e6810b3174365c4a2ade6c97ac179b3873a98d2dbc02c5819b5e9495` | | `javits-pallet-jacks.webp` | [Commons](https://commons.wikimedia.org/wiki/File:New_York_National_Guard_-_49701157361.jpg) | The National Guard | CC BY 2.0 | `60fce7ccbcbbfbda7cb1dedd452c9eb588553bbe661b6a9df44fea8003d588bc` | none | `58adb1eee44bbcac4e20ae5d4c4363e4cef4c3a906c54227c703e516c1f4d649` | | `hangar-pallet-jacks.webp` | [Commons](https://commons.wikimedia.org/wiki/File:26th_MEU_Hurricane_Sandy_Response_121104-M-SO289-034.jpg) | Sgt. Chris Stone, U.S. Marine Corps | Public domain | `4fc723406ca72ac88c0cd1f46895995953509070688c4e7695eb9df504810f2f` | none | `0e007859fee88e935665224a2e225cfcdb0881b18d0022503a4ab56a44953be9` | | not served | [Commons](https://commons.wikimedia.org/wiki/File:405th_AFSB_conducts_professional_development_staff_visit_to_Amazon_fulfillment_center_(9079637).jpg) | Cameron Porter, U.S. Army 405th Army Field Support Brigade | Public domain | `f26dbff2a929105cbbec90e667a941569953ec6569c5d8a8f69cc2898813833b` | none | `9d01fd54103e32fce51f56fb0a8f4240e8ff5671380182001d7ac3b9e1df5724` | The WebP files are the sent JPEGs re-encoded with sharp at quality 68, with the same pixel dimensions. CC BY photographs are credited on their cards and above.