Seabed - sea cucumbers

Tasks

Object detection

Outputs

bounding boxes, class names

Classes (1)

Sea Cucumber

Specifications

Training data

sea-cucumbers-new-tiles, from the RF100-VL benchmark

Accuracy

Boxes

mAP50-95

0.735

mAP50

0.932

mAP75

0.887

Precision

87.9%

Recall

88.9%

F1

88.4%

Held-out validation subset: 81 photos, 180 objects (a class-balanced 20% of the validation split). Precision, recall and F1 at confidence 0.30, IoU 0.5 · measured Nov 9, 2026

Boxes, whole benchmark

mAP50-95

0.697

RF100-VL benchmark, all 100 datasets · measured Nov 10, 2026

Validate this model yourself on /detect

Image requirements

JPG, PNG or WebP, up to 10 MB. The model works at 1280 px and expects photos like the example: it was trained on this kind of image only.

Video requirements

Photos only.

Input / output

In: one image, multipart/form-data fields file and project_id. Out: JSON detections, each with class and bbox [x1, y1, x2, y2] in pixels.

Licence

Use through the Comexp API under our Terms of service.

Version

Not stated

Hardware

Runs in the Comexp cloud: nothing to install, any device that can send an HTTP request. On your own servers or devices: on request.

API

curl -X POST https://ml.comexp.net/api/detect \
  -H "Authorization: ApiKey <YOUR_API_KEY>" \
  -F "project_id=tape_sea_cucumber" \
  -F "file=@image.jpg"

Each call costs one detection from your plan, as on /detect. Your API key is in your account settings.