FungiTastic Few-Shot Dataset

• FungiTastic Few-Shot Dataset is a specialized subset of the FungiTastic corpus focused on rare fungal species recognition using expert-curated, temporally-split data.
• It supports both classification and training-free fine-grained segmentation protocols, leveraging prototype-based inference and multimodal inputs to address low-data challenges.
• The benchmark employs geometry-aware embedding techniques and balanced evaluation metrics to overcome severe data scarcity and fine-grained recognition difficulties.

FungiTastic Few-Shot Dataset denotes the few-shot component of the broader FungiTastic fungal recognition corpus and, in related low-data segmentation work, the use of a segmentation-annotated FungiTastic subset under few-shot and few-hundred-shot protocols. In its primary classification form, it is the official rare-species subset FungiTastic–Few-shot (fungif), constructed from expert-curated fungal observations collected over roughly twenty years and organized as a temporally split benchmark for species-level recognition under severe data scarcity. In adjacent work, the same ecosystem is used to study fine-grained semantic segmentation on a segmentation-annotated subset with 194 mushroom classes, where low-shot prototype-based inference is evaluated without task-specific training (Picek et al., 2024, Cavada et al., 21 May 2026, Tam et al., 11 Jul 2025).

1. Origin within the FungiTastic corpus

FungiTastic is a multi-modal benchmark built from expert-labeled fungal observations collected by the Atlas of Danish Fungi over approximately twenty years. The full corpus comprises more than 350k observations and about 650k photographs, with additional modalities including satellite imagery, climatic time-series, location and context metadata, and segmentation resources in selected subsets. The main FungiTastic subset contains 2,829 species; FungiTastic–Mini contains 215 species; and the few-shot subset, FungiTastic–Few-shot (fungif), contains 2,427 species (Picek et al., 2024).

The few-shot subset is defined by rarity in the original corpus: all categories with less than five samples in the main FungiTastic dataset are removed from the main subset and collected into fungif. This makes fungif a benchmark for image-based few-shot classification over rare, visually similar fungal taxa rather than a conventional episodic meta-learning testbed (Picek et al., 2024).

The broader FungiTastic benchmark is explicitly multi-modal. Available data can include photographs, microscopic images, geospatial satellite patches, climatic variables, and metadata such as timestamp, GPS, habitat, substrate, toxicity, taxonomic ranks, EXIF, elevation, land cover, and biogeographical zone. Metadata are available for approximately 99.9% of observations. FungiTastic also includes DNA-based test sets for the main dataset and the Mini subset, which the dataset paper describes as providing unprecedented label reliability in those test sets; DNA confirmation for fungif is not specified (Picek et al., 2024).

Access points documented for FungiTastic include Kaggle, project documentation with full-size images and Croissant metadata, GitHub documentation, the FGVC training framework, and a Hugging Face collection of pre-trained models. Licensing is not specified in the paper (Picek et al., 2024).

2. Dataset composition, splits, and naming conventions

The official FungiTastic–Few-shot split follows a temporal protocol: training data extend up to the end of 2021, validation is drawn from 2022, and test data from 2023. The dataset paper reports the following split statistics for fungif (Picek et al., 2024).

Split	Observations / Images	Species
Train	4,293 / 7,819	2,427
Validation	1,099 / 2,285	570
Test	998 / 1,909	566

This benchmark is closed-set and does not contain an “unknown” category. The paper further notes that the few-shot dataset has no severe class imbalance like the other FungiTastic subsets, although it remains a low-data regime because all included species have fewer than five training observations in the source corpus (Picek et al., 2024).

A competition-oriented packaging appears in FungiCLEF 2025. There, the same few-shot setting is framed around category_id as the prediction target rather than raw species identifiers. The reported counts are 7,819 training images, 2,285 validation images, and 1,911 test images, with 2,413 unique species and 2,427 unique category_id values in training. The paper explicitly notes a slight mismatch between species counts and category_id counts, indicating curation steps that map species into competition classes (Tam et al., 11 Jul 2025).

This dual nomenclature is important. In the dataset paper, “FungiTastic–Few-shot” denotes a rare-species subset of the larger corpus. In the competition paper, “FungiTastic Few-Shot dataset” refers to the official challenge data bundle derived from that subset, enriched with packaged metadata and generated captions and evaluated by top-5 accuracy rather than the top-1-centered reporting used in the original dataset benchmark (Picek et al., 2024, Tam et al., 11 Jul 2025).

3. Labels, modalities, and annotation granularity

The classification benchmark is species-level and fine-grained. The underlying labels are taxonomic, and the competition target category_id is aligned to taxonomic categories but not identical to species IDs. This label space embodies the standard fine-grained fungal recognition difficulties: subtle inter-species differences and substantial intra-species variability driven by age, season, substrate, and local conditions (Tam et al., 11 Jul 2025).

For fungif, the available modalities include images, metadata, and microscopic images. The dataset paper also states that geospatial satellite data and climatic time-series are available for all observations. However, segmentation masks are not available for fungif. Human-verified body part instance masks—cap, gills, pores, ring, stem—exist for FungiTastic–Mini, where masks were generated in CVAT using SAM, but not for the main or few-shot subsets (Picek et al., 2024).

The low-data segmentation line of work uses a different portion of FungiTastic: a segmentation-annotated subset comprising approximately 13,000 training images with pixel-level segmentation annotations and 9,000 test images across 194 visually similar mushroom classes. Here the labels are species-level, such as “boletus edulis” and “amanita muscaria,” but segmentation is driven by a higher-level macro-taxonomic concept, “mushrooms,” used for class-agnostic mask extraction. In that paper, “macro-taxonomic” means a broad, high-level taxonomic category rather than a fine-grained scientific species name (Cavada et al., 21 May 2026).

This separation of annotation regimes is central. The official few-shot classification subset provides a large rare-species benchmark without segmentation masks, whereas the segmentation baseline paper studies fine-grained semantic segmentation on a smaller taxonomy with dense pixel supervision. A plausible implication is that “FungiTastic Few-Shot Dataset” in current usage denotes not a single monolithic artifact but a family of low-data evaluation settings rooted in the same fungal corpus (Picek et al., 2024, Cavada et al., 21 May 2026).

4. Evaluation protocols and metrics

The fungif benchmark is not specified as an episodic $N$-way, $K$-shot task collection. Instead, evaluation is standard closed-set classification on the temporal validation and test splits across all few-shot classes. The main metric is top-1 accuracy, and the dataset paper also reports macro-averaged $F_1$-score and top-$k$ accuracy (Picek et al., 2024).

The formulas reported in the dataset paper are:

$$F = \frac{1}{C} \sum_{c=1}^{C} F_c, \quad F_c = \frac{2 P_c \cdot R_c}{P_c + R_c}$$

and

$$\mathrm{Acc@k} = \frac{1}{N} \sum_{i=1}^{N} \mathbf{1}\left(y_i \in q_k(x_i)\right).$$

In the FungiCLEF 2025 competition setting, the official leaderboard metric is top-5 accuracy:

$$\mathrm{Top\text{-}k\ Accuracy} = \frac{1}{N} \sum_{i=1}^{N} 1[y_i \in \hat{Y}_i^{(k)}].$$

The competition paper does not describe episodic internal assessments, query-set sizes per class, or meta-learning episode construction (Tam et al., 11 Jul 2025).

The segmentation benchmark follows a different protocol. A fixed split of roughly 13k training and 9k test images across 194 classes is used. For each shot count $k$, training subsets are constructed by uniformly sampling from the training pool; because the data are long-tailed, classes with fewer than $k$ available samples contribute fewer images. Experiments are repeated across $n=20$ random seeds, and results are reported as mean metrics over those runs, with maximum standard deviation $K$0 in the tabulated results. No additional validation split is described; evaluation is reported directly on the test set (Cavada et al., 21 May 2026).

For segmentation, the paper reports mean class accuracy (mAcc) for image-level recognition and mean intersection-over-union (mIoU) for pixel-level segmentation. The paper explains these intuitively as per-class averages: IoU for class $K$1 is $K$2, and mIoU averages IoU across classes; mAcc averages the per-class image-level accuracies (Cavada et al., 21 May 2026).

5. Classification baselines on the rare-species benchmark

The original FungiTastic paper evaluates image-only few-shot baselines using both embedding-based methods and standard supervised classifiers. Among embedding methods, CLIP (ViT-B/32), DINOv2 (ViT-B/16), and BioCLIP (ViT-B/32) are paired with either 1-NN or centroid classification. BioCLIP with centroid classification is the strongest reported baseline in top-1 accuracy, reaching 21.8 top-1, $K$3, and top-3 accuracy 32.6. Its per-shot top-1 values are 12.7 for 1-shot, 23.5 for 2-shot, 32.9 for 3-shot, and 40.3 for 4-shot. DINOv2 centroid yields 17.9 top-1 and top-3 accuracy 27.8, while CLIP remains substantially weaker, with 7.2 top-1 and top-3 accuracy 13.0 under centroid classification (Picek et al., 2024).

Standard discriminative models trained with cross-entropy loss are less competitive in this setting. The reported values are 11.0 and 11.4 top-1 for BEiT-B/p16 at $K$4 and $K$5, 14.0 and 15.4 for ConvNeXt-B, and 13.9 and 19.5 for ViT-B/p16. This performance gap supports the paper’s conclusion that, under extreme rarity, frozen embedding spaces with nearest-neighbor or centroid classifiers are more effective than straightforward supervised training on the few-shot split (Picek et al., 2024).

The competition paper studies a different set of practical baselines built on pre-computed 768-D image embeddings with a linear classifier. PlantCLEF 2024 ViT-B/14 and FungiTastic ViT-B/16 both obtain 48.672% public top-5 accuracy, ahead of DINOv2 at 47.345% and FungiTastic BEiT at 42.477%. The strongest post-competition ablation reported in that paper combines PlantCLEF embeddings with Mixup at $K$6, 50 epochs, and weighted sampling, reaching 53.982 public and 46.054 private top-5 accuracy. The team’s final private ranking was 35/74, and the paper states that the approach outperformed both official baselines, BioCLIP+FAISS+Prototypes and BioCLIP+FAISS+NN (Tam et al., 11 Jul 2025).

The same paper also tests multimodal extensions by concatenating metadata-derived text and Molmo-generated captions with image features. In that pipeline, adding text decreased accuracy, and the authors attribute this to the inclusion of weakly informative fields such as district, countryCode, and hasCoordinate rather than carefully selected metadata. This suggests that multimodal supervision is not automatically beneficial in rare-species fungal FGVC; feature selection and domain adaptation matter at least as much as modality count (Tam et al., 11 Jul 2025).

6. Training-free fine-grained segmentation on the segmentation-annotated subset

The segmentation baseline paper formulates low-data fine-grained semantic segmentation as a two-stage, training-free pipeline that decouples segmentation from classification. Stage 1 uses SAM3 with the macro-taxonomic prompt “mushrooms” to obtain class-agnostic mushroom masks, thresholding output mask logits at 0.5. Stage 2 uses frozen DINOv3 image-level $K$7 embeddings of dimension 4096 and assigns fine-grained labels by prototype matching in embedding space. The predicted class label is then propagated to the pixels inside the SAM3 mask, converting class-agnostic segmentation into fine-grained segmentation without class-specific segmentation prompts (Cavada et al., 21 May 2026).

Prototype classification is defined in the paper as follows. Let $K$8 be the transformed DINOv3 embedding for image $K$9, and let $F_1$0 be the set of $F_1$1 training samples for class $F_1$2. Then

$F_1$3

with cosine similarity

$F_1$4

and decision rule

$F_1$5

The critical methodological detail is feature-space transformation. The paper applies PCA followed by whitening:

$F_1$6

followed by $F_1$7-normalization before prototype averaging. According to the authors, this “geometry correction” suppresses nuisance-dominated directions associated with backgrounds, illumination, zoom, and broader capture conditions, all of which are prominent in FungiTastic (Cavada et al., 21 May 2026).

The segmentation-only oracle comparison strongly favors macro-taxonomic prompting over fine-grained prompting. Macro prompting with “mushrooms” yields $F_1$8 with 9,643 non-empty masks out of 9,763 test images. Fine-grained prompting with oracle species names yields $F_1$9 with 6,341 non-empty masks out of 9,763 test images and requires thresholding at 0.3 because specific class prompts produce lower-confidence SAM3 outputs. This motivates the adopted macro-to-fine ordering (Cavada et al., 21 May 2026).

The few-shot and few-hundred-shot classification results further show that whitening is the decisive component. For mAcc, the “PCA white cosine” variant improves from 0.33 at $k$0 to 0.55 by $k$1, remaining at 0.55 for $k$2 and $k$3. Raw normalized cosine stays near 0.24–0.33 across the same range, and “PCA cosine” without whitening remains near 0.23–0.32. For segmentation mIoU, whitening rises from 0.18 at $k$4 to 0.31 at $k$5 and $k$6, with 0.30 at $k$7, whereas the non-whitened variants remain near 0.12–0.15. The teaser figure summarized in the paper indicates performance saturation after roughly 40–60 images per class (Cavada et al., 21 May 2026).

Computationally, the baseline is designed to scale better than class-specific prompting. A class-specific SAM3 strategy would require 194 forward passes per image, one per class, so the segmentation cost is $k$8. The macro-to-fine approach uses one SAM3 pass plus an $k$9 prototype similarity search in DINOv3 space, which keeps segmentation cost constant with respect to the number of fine-grained classes (Cavada et al., 21 May 2026).

7. Limitations, interpretation, and relation to earlier fungi few-shot benchmarks

Several limitations recur across the FungiTastic few-shot literature. For fungif, geographic bias remains a concern because most data come from Denmark, and collector behavior may overrepresent frequently visited areas. Climatic monthly time-series have missing values beyond 2020 for observations from 2020–2024. DNA-based label confirmation is specified for the main and Mini datasets, but not for fungif. The few-shot baselines are image-only, even though the dataset exposes substantial multimodal information (Picek et al., 2024).

The competition paper emphasizes additional practical difficulties: severe long-tail behavior, validation volatility for very rare classes, occasional corrupted or truncated JPEGs, and the risk that unfiltered metadata or generated captions can reduce rather than improve performance. Zero-shot pipelines based on general-purpose multimodal LLMs or VLMs perform far worse than vision-first embedding pipelines, with reported public/private top-5 accuracies ranging from 14.159/12.548 for Gemini 2.0 Flash down to 4.424/3.104 for Mistral Medium 3 (Tam et al., 11 Jul 2025).

The segmentation benchmark exposes a different bottleneck. Its main failure mode is not macro-level mask extraction but representation geometry for fine-grained discrimination. Visually similar species, strong nuisance variability, and extremely small $$F = \frac{1}{C} \sum_{c=1}^{C} F_c, \quad F_c = \frac{2 P_c \cdot R_c}{P_c + R_c}$$0 produce prototype confusion in frozen DINOv3 embeddings. Whitening helps but does not eliminate overlap for the hardest class pairs. The pipeline also assumes single-class segmentation per image by propagating one predicted label across the macro mask, and multi-class, multi-instance scenarios are not evaluated (Cavada et al., 21 May 2026).

In historical context, FungiTastic follows earlier work on fungi few-shot benchmarks but departs from their structure. The Danish Fungi 2020 few-shot benchmark proposed in 2022 used fixed episodic testbeds—5-way 1-shot, 5-way 5-shot, 100-way 1-shot, and 100-way 5-shot—with 5,000 tasks per setting and 10 query images per class. That earlier work showed a strong performance drop in many-way fine-grained regimes, with best 100-way 1-shot top-1 accuracies around 9–10%, and argued that standard few-shot benchmarks were too coarse and semantically unrealistic (Bennequin et al., 2022). FungiTastic extends the fungi-focused few-shot agenda into a larger, expert-curated, temporally split, multimodal benchmark and, in the segmentation line, into training-free low-data semantic segmentation (Picek et al., 2024, Cavada et al., 21 May 2026).

Taken together, the FungiTastic Few-Shot Dataset occupies a distinctive position in the few-shot literature. It combines real temporal drift, expert species labels, a very large rare-species taxonomy, and auxiliary modalities that remain underexploited in current baselines. The available evidence indicates that domain-adapted embeddings, careful balancing strategies, and geometry-aware prototype matching are more effective than naive supervised training or generic zero-shot prompting, while also indicating that fine-grained fungal recognition remains far from solved under realistic low-data conditions (Picek et al., 2024, Tam et al., 11 Jul 2025)