Spaces:
Running on Zero
Running on Zero
A newer version of the Gradio SDK is available: 6.28.0
metadata
title: Aerial Imagery Semantic Segmentation
emoji: 🛰️
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
short_description: Aerial semantic segmentation — CABiNet & YOLO26 models
python_version: '3.12'
startup_duration_timeout: 30m
pinned: false
UAVid Semantic Segmentation — CABiNet & YOLO26 Model Zoo
Interactive demo of the UAVid Semantic Segmentation Model Zoo — pick any model from the dropdown and run it on an oblique aerial / drone urban scene, trained on the UAVid benchmark:
| Model | mIoU (%) | Params (M) | FLOPs (GFLOPs) | HF Weights |
|---|---|---|---|---|
| CABiNet (MobileNetV3-Large) | 68.60 | 9.17 | 54.8 | link |
| CABiNet (MobileNetV3-Small) | 66.84 | 5.36 | 44.1 | link |
| YOLO26x-sem | 64.41 | 40.16 | 430.9 | link |
| YOLO26l-sem | 63.28 | 17.87 | 192.4 | link |
| YOLO26m-sem | 61.98 | 14.32 | 152.3 | link |
| YOLO26s-sem | 61.69 | 6.50 | 44.4 | link |
| YOLO26n-sem | 58.17 | 1.63 | 11.4 | link |
CABiNet (MobileNetV3-Large) is the top performer — it beats every YOLO26 variant, including the largest (YOLO26x), on mIoU while using a fraction of the compute.
Every pixel is classified into one of 8 classes: Clutter, Building, Road, Static Car, Tree, Vegetation, Human, Moving Car. The demo renders a colored overlay and a per-class legend for whichever model you select.
Example images are real UAVid validation frames from
dronefreak/UAVid-2020.
Runs on ZeroGPU.