TorchGeoDetector

class geowombat.detect.TorchGeoDetector(model='faster-rcnn', weights=None, num_classes=None, classes=None, device='auto')

Bases: GeoWombatDetector

Detection wrapper around TorchGeo / torchvision detection models.

Supports Faster R-CNN and RetinaNet via torchvision with optional pretrained weights from TorchGeo (e.g. xView). Axis-aligned only.

Parameters:

model{‘faster-rcnn’, ‘retinanet’}

Detection head. Default ‘faster-rcnn’.

weightsstr, optional

TorchGeo weights enum string (e.g. ‘FCN_RESNET50_XVIEW’) or path to a state dict. If None, uses torchvision COCO pretrained.

num_classesint, optional

Number of classes including background. Required when loading a custom-trained checkpoint.

classeslist of str, optional

Class names corresponding to non-background ids 1..N.

devicestr

‘cpu’, ‘cuda’, or ‘auto’. Default ‘auto’.

Attributes:

class_names

Methods

predict(src[, tile_size, overlap, conf, ...])

Run tiled, georeferenced inference over a raster.

Notes

For aerial / satellite imagery, the most useful TorchGeo weights are typically ‘FASTERRCNN_RESNET50_FPN_XVIEW’ (when available in your TorchGeo version). Check torchgeo.models for the current catalogue.

Attributes Summary

oriented

Attributes Documentation

oriented = False