qhjqhj00 fsl-episodic-eval
1S-DAug: One-Shot Data Augmentation for Robust Few-Shot Generalization — Bai et al. (2026) (arXiv:2602.00114, 2026)
What this evaluates
This benchmark evaluates few-shot generalization capability by measuring classification accuracy in an episodic setting where models must recognize novel classes using only a few labeled support examples. It probes the model's ability to adapt quickly to new categories and its robustness to test-time data augmentation.
Datasets
- miniImagenet — total ?; splits: train (-1), val (-1), test (-1)
- tieredImagenet — total ?; splits: train (-1), val (-1), test (-1)
- CUB — total ?; splits: train (-1), val (-1), test (-1)
- Animals — total ?; splits: train (-1), val (-1), test (-1)
Metrics
accuracy(primary) — range: percent- Percentage of correctly classified query images across sampled episodes, computed as (correct predictions / total query images) × 100. Reported as mean accuracy with 95% confidence intervals.
Input / output format
Input: Episodic few-shot setup: a support set containing K images per class (K=1 or 5) and a query set containing multiple images per class. Each instance consists of image pairs/groups passed through a backbone encoder to extract embeddings.
Output: Class label predictions for each query image, determined by nearest-neighbor matching (Euclidean distance or cosine similarity) to class prototypes in the embedding space.
Scoring recipe
def compute_accuracy(predictions, gold_labels):
correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
total = len(gold_labels)
return (correct / total) * 100
Common pitfalls
- Augmenting only the support set or only the query set causes distribution mismatch and degrades performance; both must be augmented identically.
- Removing image conditioning during diffusion generation or using full noise (η=1.0) severely compromises class faithfulness and accuracy.
- Traditional geometric test-time augmentations (rotations, affine warps) yield significantly lower accuracy (~57.9%) compared to diffusion-based methods.
Evidence (verbatim from paper)
We sample 15,000 5-way-1/5-shot queries and report mean accuracy with 95% confidence intervals across episodes. This follows the standard protocol used in related FSL work (Ye et al., [2020]; Snell et al., [2017]).
Citation
@misc{bai20261sdaug,
title={1S-DAug: One-Shot Data Augmentation for Robust Few-Shot Generalization},
author={Bai et al. (2026)},
year={2026},
note={arXiv:2602.00114}
}
- arXiv: 2602.00114