Reproduce “Low-Rank Adaption on Transformer-based Oriented Object Detector for Satellite Onboard Processing of Remote Sensing Images” 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
Xinyang Pu, Feng Xu
Researchers at Fudan University developed LoRA-Det, a hybrid parameter-efficient fine-tuning strategy for transformer-based oriented object detectors, enabling efficient model updates on satellites with limited communication bandwidth. The method achieved detection performance comparable to full fine-tuning, reaching up to 100% of its mean Average Precision on the HRSC2016 dataset, while updating only 12-14% of the total model parameters.
Key insights
- Parameter-efficient fine-tuning (PEFT) can be successfully extended to complex vision tasks like oriented object detection by strategically applying low-rank adaptation within the detector's architecture.
- A hybrid fine-tuning approach, combining LoRA for large parameter matrices with full fine-tuning for critical, less parameter-dense modules, optimizes the performance-efficiency trade-off for complex detector architectures.
- The intrinsic dimensionality of weight matrices in normal-sized models can be higher than often assumed for very large models, necessitating a data-driven approach (e.g., SVD analysis) for optimal low-rank value selection in LoRA modules.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.