Object Detector API

About the Object Detector API

Detection runs DETR ResNet-50, an object detection model from Meta AI trained on the COCO dataset's 80 everyday object classes and released under the Apache 2.0 licence. APIVerve runs it on its own servers, so images are never sent to a third-party AI provider, and photos are turned upright from their EXIF orientation before detection so boxes match the image as it is displayed.

Each object has a confidence from 0 to 1, a pixel box (x, y, width, height from the top-left corner) and the same box as fractions of the image width and height, so you can draw it at any display size. counts gives the number of each label found, ready for inventory, traffic or occupancy counts.

How it works

Detection uses DETR (DEtection TRansformer) with a ResNet-50 backbone, trained by Meta AI on COCO 2017 and released under Apache 2.0, run as an 8-bit quantized ONNX model on APIVerve's servers. The image is decoded, turned upright from its EXIF orientation and, if larger, scaled so its longer side is at most 1,333 pixels, which is the most the model reads. DETR predicts 100 candidate objects per image directly, with no anchor boxes or non-maximum suppression; candidates below minConfidence are dropped and the rest are mapped back to the original image size.

confidence is the model's class probability, 0 to 1, rounded to 4 places. box is in pixels of the upright image: x and y are the top-left corner, measured from the top-left of the image. normalizedBox divides those by imageWidth and imageHeight. Objects are sorted by confidence, and counts tallies the returned objects by label, after the labels filter.

labels accepts the 80 COCO class names (traffic light, cell phone, dining table and so on), case-insensitive, with underscores or hyphens for spaces, simple plurals and common synonyms (sofa, television, phone, bike, motorbike, fridge). An unknown label returns 400 naming it. URLs are downloaded by APIVerve from public hosts only, with a 10 MB limit; the image service never fetches caller URLs itself.

What people use it for

Counting and Occupancy

Count people, cars or bicycles in camera snapshots to measure footfall, parking use or traffic.

Image Tagging

Tag uploaded photos with the objects they contain so they can be searched and filtered in a media library.

Content Checks

Confirm that listing photos actually show the item for sale, such as a car, bicycle or couch, before they are published.

Smart Cropping

Crop images around the main subject using its bounding box instead of the centre of the frame.

Ways to call it

One endpoint, many ways in — REST with JSON, XML, YAML and CSV, plus GraphQL and an MCP interface for AI agents.

JSON

Default REST response

XML

Markup format

YAML

Human-readable

CSV

Tabular export

Beta

GraphQL

Query language

New

MCP

For AI agents

Request parameters — POST /v1/objectdetector

Sent as JSON in the request body. Premium parameters are accepted on every plan but only take effect on plans that include them. Anything not listed here is dropped rather than passed through.

Optional

1
2
3
4
5

curl -X POST "https://api.apiverve.com/v1/objectdetector" \
  -H "x-api-key: YOUR_API_KEY" \
  -F "image=@photo.jpg" \
  -F "minConfidence=0.5" \
  -F "labels=person,car"

Authentication

Send your key in the x-api-key header. That is the only auth step — no token exchange, and no per-endpoint scope to configure.

Response — POST /v1/objectdetector

Every APIVerve endpoint returns the same three top-level keys, so one response handler covers your whole integration: status, error and data. Only data changes shape.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98

{
  "status": "ok",
  "error": null,
  "data": {
    "imageWidth": 640,
    "imageHeight": 480,
    "objectCount": 5,
    "counts": {
      "cat": 2,
      "remote": 2,
      "couch": 1
    },
    "objects": [
      {
        "label": "cat",
        "confidence": 0.999,
        "box": {
          "x": 4,
          "y": 51,
          "width": 320,
          "height": 417
        },
        "normalizedBox": {
          "x": 0.0063,
          "y": 0.1062,
          "width": 0.5,
          "height": 0.8688
        }
      },
      {
        "label": "cat",
        "confidence": 0.9987,
        "box": {
          "x": 331,
          "y": 20,
          "width": 309,
          "height": 350
        },
        "normalizedBox": {
          "x": 0.5172,
          "y": 0.0417,
          "width": 0.4828,
          "height": 0.7292
        }
      },
      {
        "label": "remote",
        "confidence": 0.9984,
        "box": {
          "x": 329,
          "y": 67,
          "width": 42,
          "height": 128
        },
        "normalizedBox": {
          "x": 0.5141,
          "y": 0.1396,
          "width": 0.0656,
          "height": 0.2667
        }
      },
      {
        "label": "remote",
        "confidence": 0.9975,
        "box": {
          "x": 29,
          "y": 68,
          "width": 160,
          "height": 50
        },
        "normalizedBox": {
          "x": 0.0453,
          "y": 0.1417,
          "width": 0.25,
          "height": 0.1042
        }
      },
      {
        "label": "couch",
        "confidence": 0.9961,
        "box": {
          "x": 0,
          "y": 0,
          "width": 636,
          "height": 472
        },
        "normalizedBox": {
          "x": 0,
          "y": 0,
          "width": 0.9938,
          "height": 0.9833
        }
      }
    ],
    "minConfidence": 0.5,
    "labels": null
  }
}

Response fields

Paths are relative to data. Premium fields are absent rather than zeroed on plans that do not include them, so check for presence instead of comparing to 0.

Response headers

Every response carries your balance, so your own code always knows where it stands without polling anything. All four are exposed to browsers through CORS.

Errors

Read the HTTP status first, then error for the specific reason. The body names the parameter that has to change.

What a call costs — 10 credits per call

Credits are shared across every APIVerve API — one balance, one key, one bill. Failed requests never count against it, so a rejected input costs nothing.

Calls are what the credits buy at this API's rate — spend them here, on any of the other APIs, or across both. Nothing is reserved per endpoint.

Zero-code embed — one snippet, no backend

Drop an interactive Object Detector form onto any page. The widget calls the API for you, so no key ever appears in your markup and there is nothing to deploy.

Changelog — Object Detector API

Every change to this endpoint, newest first. Breaking changes ship as a new version and the previous one keeps serving — response fields are added, never removed or retyped in place, so an integration written against v1 keeps working.

Currently v1