AGenDA-Synthetic-UGRC

Introduction

After fine-tuning Stable Diffusion, we generate synthetic source domain data (LINZ) and target domain data (UGRC). For each synthetic data sample, we extract the cross-attention maps from the word "car", the foreground learnable token, and the background learnable token during denoising steps. The enhanced cross-attention map is then obtained by stacking these cross-attention maps. First, we can label synthetic source domain data using the trained detectors on real source domain data. Then, we can train another detector on synthetic source domain cross-attention maps and then label synthetic target domain data. Finally, we train a detector on synthetic target domain data and test on real data to evaluate our pseudo-labels.

Model Usage

This folder contains four detectors trained on our generated Synthetic UGRC data and tested on Real UGRC 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
➡️ Code: AGenDA
➡️ Synthetic Data: AGenDA