---
title: 'FungiCLEF 2025: Few-Shot Fungal Recognition Challenge'
url: https://www.emergentmind.com/topics/fungiclef-2025
type: topic
---

# FungiCLEF 2025: Few-Shot Fungal Recognition Challenge

FungiCLEF 2025 is a few-shot fungal species recognition challenge built on real field observations from the Atlas of Danish Fungi and using the FungiTastic Few-Shot subset of the FungiTastic benchmark, with the stated aim of automatic fungal species recognition under realistic, field-collected conditions [2509.11020; 2507.08248]. Within the broader FungiTastic program, which introduced about 350k multimodal observations of 6k fine-grained categories and multiple task configurations, FungiCLEF 2025 represents the few-shot, long-tail identification setting most directly aligned with practical biodiversity monitoring [2408.13632].

## 1. Benchmark basis and dataset lineage

FungiCLEF 2025 is described in challenge reports as a few-shot, fine-grained fungi classification task derived from the FungiTastic dataset and centered on the FungiTastic Few-Shot subset [2507.08248]. FungiTastic itself was introduced as a multi-modal dataset and benchmark for image categorization based on fungal records continuously collected over a twenty-year span, with support for standard closed-set classification, open-set classification, multi-modal classification, few-shot learning, domain shift, and segmentation [2408.13632].

The underlying source is the Atlas of Danish Fungi. In the FungiTastic benchmark design, the temporal partition is chronological: training uses observations up to and including 2021, validation uses 2022, and test uses 2023 [2408.13632]. This suggests that the challenge setting is not merely a random subsampling problem, but one that is naturally exposed to temporal drift, seasonal effects, and changing observation conditions.

The broader FungiTastic resource includes not only the few-shot split used for the 2025 challenge, but also full closed-set, open-set, and segmentation-oriented subsets, as well as DNA test subsets with “DNA-sequenced ground truth of unprecedented label reliability” [2408.13632]. That larger benchmark context is important because it locates FungiCLEF 2025 within a wider program of fungal computer vision rather than as an isolated leaderboard task.

## 2. Observation structure and few-shot regime

The challenge data are observation-centric rather than image-centric. Each observation typically contains several field photographs of the same specimen, with multiple resolutions including 300, 500, 700 px and full size, together with rich metadata such as time, GPS coordinates, substrate, habitat, and toxicity; additional environmental context is available through satellite imagery and weather variables, and machine-generated image captions were also released in a separate folder after the competition started [2509.11020].

Challenge reports describe the few-shot subset with slightly different tabulations. One report gives a training set of 4,293 observations, 7,819 images, and 2,427 species/classes, and a validation set of 1,099 observations, 2,285 images, and 570 classes [2509.11020]. Another reports 7,819 training images, 2,413 species, and 2,427 unique `category_id`, together with 2,285 validation images, 569 species, 570 unique `category_id`, and 1,911 test images with labels withheld [2507.08248]. Both reports agree on the defining structural property: the training split is extreme few-shot and long-tailed.

That long-tail is explicit. One report states that each species has 1–4 observations in train, so “almost every class is few-shot” [2509.11020]. The same report also notes that, according to the organizers, about 20% of all verified fungi observations, representing nearly 20,000 instances, are associated with rarely recorded species [2509.11020]. Another report emphasizes that the most frequent class has about 30 images, whereas many classes have only 1 image, and that the training and validation splits have different class distributions [2507.08248].

A common misconception is that FungiCLEF 2025 is simply a large closed-set species benchmark with sufficient supervision. The reported data regime indicates the opposite: it is a closed-set evaluation problem built on a severe few-shot and imbalance structure, where category coverage is large but per-class evidence is minimal [2509.11020; 2507.08248].

## 3. Evaluation protocol and benchmark logic

The main evaluation criterion is top-k retrieval of the correct species. One competition report defines the official metric as Recall@5,
$$
\mathrm{Recall@}k = \frac{1}{N} \sum_{i=1}^{N} \mathbb{I}\left[ y_i \in \hat{Y}_i^k \right],
$$
where \(N\) is the number of test observations, \(y_i\) is the true species, and \(\hat{Y}_i^k\) is the set of top-\(k\) predicted species for observation \(i\) [2509.11020]. Another report describes the official metric as Top-k accuracy with \(k=5\),
$$
\text{Top-k Accuracy} = \frac{1}{N} \sum_{i=1}^{N} 1\left[y_i \in \hat{Y}_i^{(k)}\right].
$$
[2507.08248]

Leaderboard evaluation is split into a public leaderboard and a private leaderboard. In the former, ranking is based on a subset of the test data; in the latter, ranking is based on a hidden subset used for final evaluation [2509.11020]. Predictions are required to come from the fixed species set defined by the challenge, and one report explicitly characterizes the evaluation setting as closed-set even though many species are truly few-shot [2509.11020].

The closed-set emphasis also distinguishes FungiCLEF 2025 from earlier fungal challenge formulations. One report notes that, unlike previous FungiCLEF editions such as FungiCLEF 2022, which emphasized open-set recognition, FungiCLEF 2025 focuses on closed-set identification [2509.11020]. At the same time, the underlying FungiTastic benchmark remains broader: it also specifies open-set protocols, cost-sensitive evaluation, multi-modal classification, domain shift, and segmentation [2408.13632].

## 4. Submitted systems and reported performance

A prominent line of FungiCLEF 2025 systems uses prototype-based metric learning. In the prototypical formulation, a class prototype is the mean of support embeddings,
$$
c_k = \frac{1}{|S_k|} \sum_{(x_i, y_i) \in S_k} f_\theta(x_i),
$$
with squared Euclidean distance used during episodic training,
$$
d(x_q, c_k) = \| f_\theta(x_q) - c_k \|^2,
$$
and cosine similarity used at inference,
$$
\mathrm{sim}(x_q, c_k)
= \frac{f_\theta(x_q)^\top c_k}{\|f_\theta(x_q)\| \, \|c_k\|}.
$$
The strongest reported prototype system used a BioCLIP backbone with Prototypical Networks, \(K=75\), \(S=3\), \(Q=1\), and train+validation support [2509.11020].

Reported public and private leaderboard scores for representative BioCLIP-based systems are summarized below.

| Method | PB | PR |
|---|---:|---:|
| BioCLIP + FAISS + Prototypes (organizers) | 0.33185 | 0.26649 |
| BioCLIP prototype averaging | 0.6017 | 0.5472 |
| BioCLIP ProtoNet, \(K=70, S=3, Q=1\), train only | 0.61061 | 0.57179 |
| BioCLIP ProtoNet, \(K=75, S=3, Q=1\), train + val | 0.64159 | 0.58473 |

These numbers show a large gap between the official FAISS-based baseline and prototype learning on top of the same general family of embeddings [2509.11020]. The same report states that the best Prototypical Network exceeds the competition baseline by more than 30 percentage points in Recall@5 on both public and private leaderboards [2509.11020].

A second methodological line, reported by team DS@GT, used frozen vision transformers as feature extractors followed by a shallow classifier. Four backbones were evaluated in feature-extractor mode: DINOv2 ViT-B/14, a PlantCLEF 2024 ViT-B/14 model, FungiTastic BEiT-B/16, and FungiTastic ViT-B/16 [2507.08248]. On public top-5 accuracy, PlantCLEF 2024 and FungiTastic ViT both achieved 48.67%, ahead of DINOv2 at 47.35% and FungiTastic BEiT at 42.48% [2507.08248]. Their best post-competition private result was 47.35% with PlantCLEF 2024 embeddings and feature-level mixup with \(\alpha = 1.45\), while their competition submission achieved 49.56% public and 45.41% private, ranking 37/74 on the public leaderboard and 35/74 on the private leaderboard [2507.08248].

## 5. Methodological findings

The most consistent methodological finding across FungiCLEF 2025 reports is the importance of domain-adapted visual representations. In the prototype-based study, BioCLIP outperformed larger generic models such as ViT-H/14, and the authors explicitly interpret this as evidence that appropriate pretraining is more valuable than sheer model size in this setting [2509.11020]. In the DS@GT report, PlantCLEF 2024 and FungiTastic-specific backbones slightly outperformed generic DINOv2, again pointing toward the value of domain-adapted transfer [2507.08248].

A second finding concerns the efficacy of episodic few-shot training. The best prototype systems used high-way episodes, pushing \(K\) as high as GPU memory allowed, and one report argues that large-way episodes better match the true large-class scenario of the challenge [2509.11020]. The same report also found that including both train and validation observations in the support set improved robustness of class prototypes [2509.11020].

A third finding concerns regularization under extreme imbalance. DS@GT applied mixup directly in embedding space:
$$
\tilde{x} = \lambda x_i + (1 - \lambda) x_j, \qquad
\lambda \sim \text{Beta}(\alpha, \alpha),
$$
with the corresponding loss
$$
\mathcal{L}_{\text{Mixup}}
= \lambda \cdot \mathcal{L}\big(f(\tilde{x}), y_i\big)
+ (1 - \lambda) \cdot \mathcal{L}\big(f(\tilde{x}), y_j\big).
$$
They report that mixup was the primary driver of improvement, whereas weighted sampling gave a smaller secondary gain [2507.08248].

The challenge literature also counters several recurrent assumptions. One is that metadata and captions are automatically useful. In the DS@GT report, unfiltered metadata and text embeddings degraded performance relative to image-only baselines, and the authors attribute this to noisy or weakly informative metadata fields [2507.08248]. Another is that zero-shot LLM or VLM prompting is already competitive with vision backbones. The same study evaluated Gemini 2.0 Flash, Gemini 2.5 Flash, GPT-4.1-mini, and Mistral Medium 3 with hierarchical prompting, but all were far below the best vision-based systems, with private top-5 accuracies of 12.55%, 13.58%, 6.21%, and 3.10%, respectively [2507.08248].

## 6. Scope, limitations, and relation to adjacent fungal AI

The best-performing reported FungiCLEF 2025 systems are image-centric. One prototype report explicitly notes that, although the dataset includes rich metadata, satellite imagery, and weather data, the final best model used images only [2509.11020]. The same paper also identifies several unresolved issues: scalability of \(K\) in episodic training is constrained by GPU memory; there is no explicit long-tail loss or class reweighting in the final prototype system; and there is no open-set handling, even though out-of-distribution species remain a major practical problem [2509.11020].

These limitations are best understood against the background of the broader FungiTastic benchmark, which already supports open-set classification, multi-modal classification, domain shift, cost-sensitive evaluation, and segmentation, and which provides DNA-based test subsets intended to improve label reliability [2408.13632]. This suggests that FungiCLEF 2025 should be read as a focused few-shot instantiation rather than as a full statement of the fungal recognition problem.

The challenge also differs sharply from earlier fungal vision settings. In clinical microscopy, for example, “Deep learning approach to describe and classify fungi microscopic images” addressed 180 images across 9 strains of clinically relevant yeasts and yeast-like fungi in a controlled laboratory pipeline, with the best configuration based on AlexNet features, Fisher Vectors, and a linear SVM [2005.11772]. In unsupervised fungal image organization, “Simple but Effective Unsupervised Classification for Specified Domain Images: A Case Study on Fungi Images” reported 94.1% and 96.7% on public and private fungal datasets using ConvNeXt, UMAP, and multi-clustering with voting [2311.08995]. By contrast, FungiCLEF 2025 emphasizes realistic field observations, thousands of categories, and an extreme few-shot long tail.

Taken together, the reported evidence positions FungiCLEF 2025 as a benchmark for fine-grained, field-collected fungal recognition under realistic data scarcity. Its central technical lesson is not merely that fungal classification is difficult, but that the decisive factors are prototype quality, domain-specific pretraining, and training procedures that explicitly match the few-shot structure of the task [2509.11020; 2507.08248].

Source: https://www.emergentmind.com/topics/fungiclef-2025