AGenDA
Collection
Checkpoints for the paper "Adapting Vehicle Detectors for Aerial Imagery to Unseen Domains with Weak Supervision" (https://arxiv.org/abs/2507.20976).
•
8 items
•
Updated
•
1
Introduction
We introduce a real-world aerial view dataset, LINZ, captured in Selwyn (New Zealand). The dataset has ground sampling distance (GSD) of 12.5 cm per px and has been sampled to 112 px × 112 px image size. For data annotation, we label only the small vehicle centers. To leverage the abundance of bounding box-based open-source object detection frameworks, we define a fixed-size ground truth bounding box of 42.36 px × 42.36 px centered at each vehicle. Annotations are provided in COCO format [x, y, w, h], where "small" in the annotation json files denotes the small vehicle class and (x, y) denotes the top-left corner of the bounding box. We use AP50 as the evaluation metric.
Model Usage
This folder contains four detectors trained on Real LINZ data and tested on Real LINZ data, along with configuration files we use for training and testing.
References
➡️ Paper: Adapting Vehicle Detectors for Aerial Imagery to Unseen Domains with Weak Supervision
➡️ Project Page: Webpage
➡️ Data: AGenDA