hawk_eye.core.detector¶

Overview¶

UAV uses object detectors to predict the shape type of the AUVSI SUAS targets. Currently, the detectors are defined by the RetinaNet architecrure with a backbone, FPN, RetinaNet head, and anchor box regression. This architecture works accross a variety of tasks, including ours, but the architecture can be daunting to beginners.

Newer models like FCOS, CenterNet, and Foveabox, are a class of anchor-less detectors which are simpler architectures. These models would likely work well on our project as well and are worth considering as alternatives.

For UAV, the targets are composed of a shape and alphanumeric (A-Z or 0-9). Due to the limitations of our camera, we somtimes lack the resolution for a model to determine the alphanumeric (even a human sometimes can’t). This is the motivating factor for why the object detectors we run only predict the shape class of the target.

Module¶

Todo

Write function docs.

A detector model which wraps around a feature extraction backbone, fpn, and RetinaNet head.This allows for easy interchangeability during experimentation and a reliable way to load saved models.

class hawk_eye.core.detector.Detector(model_params: dict = None, timestamp: str = None, confidence: float = 0.05, num_detections_per_image: int = 100, half_precision: bool = False)[source]¶: Initializes internal Module state, shared by both nn.Module and ScriptModule.