Reproduce “Adapting Vehicle Detectors for Aerial Imagery to Unseen Domains with Weak Supervision” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiao Fang, Minhyek Jeon, Zheyang Qin, Stanislav Panev, Celso de Melo, Shuowen Hu, Shayok Chakraborty, Fernando De la Torre

AGenDA uses weak target vehicle-presence labels to create target-like synthetic aerial images, then labels them using diffusion-model attention maps. It trains a detector through transferred pseudo-labels, adapting from fully annotated source imagery without requiring target vehicle boxes. On LINZ-to-UGRC, Faster R-CNN reaches 69.3% AP50, versus 63.1% for OCUD; results cover small-vehicle locations in the tested transfers.

Key insights

  • Off-the-shelf generative AI models and VLLMs struggle with geographical understanding and small object detection/generation in aerial views, requiring domain-specific fine-tuning.
  • Multi-channel cross-attention maps, incorporating learnable tokens for object and background concepts, provide more robust and precise object localization than single-channel maps.
  • A two-stage fine-tuning process for the diffusion model, where the U-Net is aligned first and then specialized tokens are learned, significantly improves adaptation performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.