EO_OS Object Detection

Version: 1

Microsoft•Last updated May 2026

Advanced object detection for Earth Observation and Overhead Sensing imagery.

Vision

Understanding

Agents

The EO/OS Object Detection Model is a specialized computer vision model designed for Earth Observation (EO) and Overhead Sensing (OS) applications. It excels at detecting and localizing objects in satellite and aerial imagery, providing accurate bounding boxes and classifications for targets of interest in overhead perspectives. The model is optimized for batch processing workflows, enabling efficient analysis of large-scale imagery datasets.

API Usage

The EO_OS Object Detection model exposes two endpoints for inference:

Binary Image Upload (Primary Endpoint)

Send raw image binary data directly to the root endpoint:

curl -X POST "$ENDPOINT_URL/" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: image/png" \
  --data-binary @image.png

Base64 JSON Input (HuggingFace-compatible)

Send base64-encoded image data as JSON to the /score endpoint:

# Encode image to base64
IMAGE_BASE64=$(base64 -w 0 image.png)

# Send request (threshold defaults to per-class config values)
curl -X POST "$ENDPOINT_URL/score" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"inputs\": \"$IMAGE_BASE64\"}"

# Send request with optional threshold parameter to override defaults
curl -X POST "$ENDPOINT_URL/score" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"inputs\": \"$IMAGE_BASE64\", \"parameters\": {\"threshold\": 0.3}}"

Parameters

Parameter	Type	Required	Description
`inputs`	string	Yes	Base64-encoded image data
`parameters.threshold`	float	No	Confidence threshold (0-1). If omitted, uses per-class defaults

Response Format

Both endpoints return a list of detected objects in HuggingFace format:

[
  {
    "label": "airplane",
    "score": 0.95,
    "box": {
      "xmin": 100.5,
      "xmax": 200.7,
      "ymin": 50.2,
      "ymax": 150.3
    }
  }
]

Health Check

Check service health with a GET request:

curl "$ENDPOINT_URL/health"

Evaluation Overview

Model Info

A Mask R-CNN detector using a HorNet_Tiny backbone with an FPN neck, followed by an RPN and Standard RoIHead (2FC bbox head). Trained on the DIOR (Detection in Optical Remote sensing images) dataset. For more information on using this model from Azure AI Foundry, including a client SDK to simplify working with the deployed model, see the GeoAI SDK .

Model Benchmarks

The model has been evaluated on the DIOR validation dataset.

Performance Summary

Metric	Value
mAP (AP@[0.50:0.95], maxDets=100)	0.546
AP50	0.782
AP75	0.602
AP_S	0.190
AP_M	0.456
AP_L	0.729
AR@100	0.609
AR@300	0.609
AR@1000	0.609
AR_S	0.264
AR_M	0.521
AR_L	0.789

Protocol

Metric: COCO bbox metrics
IoU thresholds: 0.50:0.95 (step 0.05), plus AP50/AP75
Area ranges: small / medium / large (COCO definitions)

Object Categories

DIOR (20 classes)

airplane
airport
baseballfield
basketballcourt
bridge
chimney
dam
Expressway-Service-area
Expressway-toll-station
golffield
groundtrackfield
harbor
overpass
ship
stadium
storagetank
tenniscourt
trainstation
vehicle
windmill

Model Specifications

LicenseCustom

Last UpdatedMay 2026

Input TypeImage

Output TypeText

ProviderMicrosoft

Quick Start