Detection Overlay

Preview in your theme

status='ready' — 12 detections, 5 classes, one colour slot each. Hover a box and its rail row lights up; hover a row and its box does. Tab lands on the list once (roving tab index), arrow keys walk it, Enter pins a detection and Escape unpins it. Drag the confidence gate: 3 detections sit below 0.5 and are already hidden. Click a class chip to switch that class off.

9 / 12

Confidence ≥

50%

Street-level photograph the detector ran on

rt-detr-l · coco-80218 ms

threshold=0.50 · selectedId=null

status='ready' with 40 overlapping boxes. Big boxes are painted first, so the small ones on top of them stay hoverable, and everything that isn't the active detection drops to 35% opacity instead of turning into a plaid. The gate starts at 0.6; the rail scrolls past its 26rem cap and the counter in the control bar is always the truth about how many survived the filters.

25 / 40

Confidence ≥

60%

rt-detr-l · coco-80441 ms

view='image' with a forced square frame — the rail is gone, so the BOXES become the listbox: they take the roving tab index, announce themselves as options and answer the same arrow keys. Exactly one surface is ever in the accessibility tree, so a screen reader never hears the same detection twice. aspectRatio=1 letterboxes a 16:10 photo, and every box lands on the PICTURE rather than on the frame.

10 / 12

Confidence ≥

40%

view='list' — no picture at all: a narrow column, a text-first report, a printed page. Same contract, same filters, same keyboard. The class swatch and the confidence bar carry what the boxes were carrying.

12 / 12

Confidence ≥

30%

rt-detr-l · coco-80218 ms

status='ready' with a picture that 404s. The detections are perfectly good data, so this is NOT the error envelope: the frame keeps the image's aspect ratio, the boxes stay drawn where they belong, the failure says so in a corner with a Reload that remounts the image element (reassigning the same src would reuse the cached failure), and the rail never notices.

9 / 12

Confidence ≥

50%

rt-detr-l · coco-80218 ms

Hostile payload over a portrait frame: a duplicate id (dropped), a box 60% wide starting at x=0.78 (trimmed at the edge), a zero-height box (floored so it stays clickable), confidence 1.4 and −0.3 (clamped, so the −0.3 one is gone at any gate above 0), a 75-character label (truncated in the chip, whole in the rail) and 6 classes over a 5-slot palette (wraps by position, never by hash). The container is only 28rem wide, so the split collapses to a stack — a container query, not a viewport one — and a 2:3 photo is letterboxed into a 3:4 frame with the boxes still on the picture.

5 / 6

Confidence ≥

20%

internal-warehouse-v3

status='ready', gate at 0.99 — nothing survives the filter. This is deliberately NOT the empty state: the model did find things, your filter hid them, so the wording names the cause and hands back a Clear filters button instead of claiming the image was blank.

0 / 12

Confidence ≥

99%

Everything is filtered out. 12 detections sit below 99% or belong to a class you switched off.

rt-detr-l · coco-80218 ms

status='loading' — inference is still out. The skeleton reserves the frame at the image's own aspect ratio, so nothing jumps when the result lands.

status='empty' — the model ran and found nothing. The picture stays on screen: that is the evidence for the sentence.

No objects were detected in this image.

status='error' — the request itself failed. A different failure from 'no objects', so it reads differently and offers a retry.

Detection failed

The inference endpoint returned 503 after 2 retries. request-id: 6c31-a0f2

Loading preview…

This component is part of Pro

The install command, prompt and concept map are included with a Pro license — one purchase unlocks every Pro component, forever.

Unlock with Pro