Lora Fewshot Aerial Eval

lora-fewshot-aerial-eval

Analyzing the Impact of Low-Rank Adaptation for Cross-Domain Few-Shot Object Detection in Aerial Images — Talaoubrid et al. (2025) (arXiv:2504.06330, 2025)

What this evaluates

Evaluates parameter-efficient fine-tuning (LoRA) for cross-domain few-shot object detection on aerial imagery. It probes the model's ability to generalize to new domains with limited labeled data while mitigating overfitting.

Datasets

• DOTA — total ?; splits: train (-1); HF HichTala/dota
• DIOR — total ?; splits: train (-1); HF HichTala/dior

Metrics

• mAP@0.5 (primary) — range: percent
  • Mean Average Precision computed at an Intersection over Union (IoU) threshold of 0.5, following the standard COCO evaluation protocol.

Input / output format

Input: RGB aerial images with bounding box annotations converted to COCO format.

Output: Predicted bounding boxes and class labels per image, formatted as COCO detection results for pycocotools.

Scoring recipe

import numpy as np
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
coco_gt = COCO('ground_truth.json')
coco_dt = coco_gt.loadRes('predictions.json')
coco_eval = COCOeval(coco_gt, coco_dt, 'bbox')
coco_eval.params.iouThrs = np.array([0.5])
coco_eval.params.maxDets = [300]
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
mAP = coco_eval.stats[0]

Common pitfalls

• Few-shot training splits are randomly sampled per class; results must be averaged over 5 independent runs to account for selection variance.
• The maximum number of detections per image is explicitly capped at 300, deviating from the default COCO limit of 100.
• LoRA is applied either directly to the pre-trained weights or to the best checkpoint from an intermediate full fine-tuning phase, requiring careful checkpoint selection.

Evidence (verbatim from paper)

We evaluated the model performance using the mean average precision (mAP) at an IoU threshold of 0.5, which is a standard metric for object detection tasks. Given that DOTA images often contain more than 100 objects, we set the maximum detection threshold to 300 in the pycocoapi evaluation toolkit [[25]].

Citation

@misc{talaoubrid2025lora,
  title={Analyzing the Impact of Low-Rank Adaptation for Cross-Domain Few-Shot Object Detection in Aerial Images},
  author={Talaoubrid et al. (2025)},
  year={2025},
  note={arXiv:2504.06330}
}

• arXiv: 2504.06330