Analyzing the Impact of Low-Rank Adaptation for Cross-Domain Few-Shot Object Detection in Aerial Images
Hicham Talaoubrid, Anissa Mokraoui, Ismail Ben Ayed, Axel Prouvost, Sonimith Hang, Monit Korn, Rémi Harvey
TL;DR
This paper addresses cross-domain few-shot object detection in aerial imagery by integrating Low-Rank Adaptation (LoRA) into DiffusionDet and evaluating on COCO-to-aerial transfers (DOTA and DIOR). It compares baseline fine-tuning, direct LoRA, and LoRA after intermediate fine-tuning across multiple ranks, finding that LoRA after intermediate fine-tuning offers modest gains in 1- and 5-shot scenarios, while full fine-tuning remains superior with more data. The study highlights LoRA as a viable, parameter-efficient adaptation strategy for resource-constrained aerial detection tasks and suggests avenues for combining LoRA with other few-shot techniques. The work provides empirical insights into when and how LoRA can balance model adaptation and generalization across challenging domain shifts, contributing to practical cross-domain FSOD strategies. $mAP$ at $IoU=0.5$ was used to assess performance, demonstrating the method's effectiveness under realistic aerial-detection metrics.
Abstract
This paper investigates the application of Low-Rank Adaptation (LoRA) to small models for cross-domain few-shot object detection in aerial images. Originally designed for large-scale models, LoRA helps mitigate overfitting, making it a promising approach for resource-constrained settings. We integrate LoRA into DiffusionDet, and evaluate its performance on the DOTA and DIOR datasets. Our results show that LoRA applied after an initial fine-tuning slightly improves performance in low-shot settings (e.g., 1-shot and 5-shot), while full fine-tuning remains more effective in higher-shot configurations. These findings highlight LoRA's potential for efficient adaptation in aerial object detection, encouraging further research into parameter-efficient fine-tuning strategies for few-shot learning. Our code is available here: https://github.com/HichTala/LoRA-DiffusionDet.