The FungiCLEF 2025 challenge encourages improvement on few-shot fine-grained classification with large scale (2,427 classes). This represents a real-world application with a dataset of rare species in Danish Fungi. In this paper, we present our approach to the challenge, which aims to classify images of fungi given few example images per species. This method utilizes pretrained embedding models DINOv2, BEIT, and SAM. Simple image augmentations are applied at both train and test time. Embeddings from each model are concatenated into a single embedding along the feature dimension per augmented version of the image. A simple projection network was trained to improve the discriminative performance of the embeddings on the training samples. Cosine similarity between the class centroid and the observation centroid is used for class prediction, as in Prototypical Networks. Finally, an ensemble of these pipelines is utilized to further boost performance. Image augmentation is shown to be the largest contributor to the performance of the solution, followed by learning an embedding projection, and utilizing multiple embedding models. Our method secured 1st place in the FungiCLEF 2025 competition on the private leaderboard. Code is available at https://github.com/Jack-Etheredge/fungiclef2025.
https://github.com/Jack-Etheredge (J. N. Etheredge)
CEUR
ceur-ws.org
5-way, 5-shot). ImageNet [
Last year’s FungiCLEF 2024 challenge [
The FungiCLEF 2025 challenge [
An observation refers to a real-world occurrence of fungi, which may include, but is not limited to, an individual mushroom, a cluster of mushrooms, or mold growing on a surface, either in a natural environment or as a collected sample. Each observation comprises one or more instances. An instance is an individual data point associated with an observation and consists of an image, its associated metadata, and a generated caption. For example, an individual mushroom may constitute an observation, but multiple images of this mushroom might be captured from diferent angles. Each of these images (along with its metadata and caption) would represent a distinct instance linked to the same observation. The solution proposed in this paper only utilizes the images, since initial experiments with captions and metadata were not promising (data not shown).
The dataset contained 2,427 classes with 5,392 observations comprising 10,104 instances between the training and validation sets. Most classes have a single observation and most observations have a single instance. All classes had fewer than 5 observations. Combining the training and validation sets into a single dataset, the class with the most instances has 39 instances. Though not as extreme as the parent FungiTastic dataset, the challenge dataset still exhibits severe class imbalance, with most classes having only a single observation while the largest class by instance count has 39 instances, creating a long-tailed distribution.
The objective of FungiCLEF 2025 was to achieve the best average performance predicting the class of each test observation given one or more instances per observation. The public and private leaderboards for the competition both used average recall at rank = 5 (recall@5), which we refer to as Top-5 accuracy or simply Top-5 hereafter.
For each test observation , let denote its true class label, and { 1̂, 2̂, … , 5̂} be the top 5 predicted classes. The recall@5 for observation is defined as:
The average recall@5 over the entire test set is then computed as: 1 if ∈ { 1̂, 2̂, … , 5̂} 0 otherwise | | ∈ (1) (2)
The overall solution is illustrated in Figure 1. Figure 1A shows that training and test time difer only by the level of hierarchy that the embeddings are averaged to. For creating the prototype embeddings, all augmented versions of images belonging to each class are averaged together. Since the competition expects observation-level predictions, all instances belonging to each test observation and all augmentations of the instance images are averaged into a single embedding. For training of the projection network, all the augmented versions of the training images are used with their class labels as targets. Predictions are made by calculating the cosine similarity between the class-level prototype embeddings and the observation-level test embeddings. Hereafter, this series of functions to transform a collection of training images into prototype embeddings and test images into observation embeddings to produce class-wise cosine similarities through the use of a specific combination of image augmentations, frozen embedding models, and a trained projection network will be referred to as an embedding pipeline. outputs along the embedding dimension, followed by a projection via a multilayer perceptron, and parameterized by . Let denote the support set for class . The class prototype p ∈ ℝ is defined as: number of augmented images in the observation:
Each observation consists of a set of images ℐ = { 1, 2, … , }, and each image has a set of augmentations ( ) = { (0) , (1), … , ( )}, where
(0) is the original image. Let denote the total
The final observation embedding z ∈ ℝ is computed as the mean of all augmented instance image embeddings belonging to the observation:
Per embedding pipeline, the embeddings used for the prototype embeddings and the test observation embeddings were projected using the same trained projection network. An ensemble of 5 embedding pipelines was used to generate the final predictions. These embedding pipelines difered only by the training-validation split and initialization of the projection network. The validation portion was used for early stopping during training of the projection network. The softmax probabilities over the classes for each model were generated from the cosine similarities between each test observation and the prototype embedding for each class. The ensemble average softmax probability of the cosine similarities was used to rank the classes per observation. The top 10 classes were returned per observation as was p = 1 | | ∑ ( ) ∈ = ∑ | ( )| ∈ℐ z = 1
∑ ∈ℐ () ∑ ∈ ( ) ( () ) (3) (4) (5) expected of the participants. Only the first 5 of these 10 classes factored into the leaderboard ranking, however, since the competition evaluation metric was the recall at top-5.
The same image augmentations were performed for both the training samples and test time augmentations. This was done both for simplicity and also to maximize agreement between the prototypes and the test embeddings. Only geometric augmentations were used in the winning solution. The specific augmentations utilized were: 80% center crop, 80% top left crop, 80% top right crop, 80% bottom left crop, 80% bottom right crop, horizontal flip, 90-degree rotation, 270-degree rotation, 15-degree rotation, and 345-degree rotation.
All experiments were performed on a machine with a single NVIDIA RTX 3090 graphics card and all
models were trained using PyTorch [
For all models, after the geometric augmentations were performed, the augmented image was resized
with bicubic interpolation to 1.14x the final image size used for that model and then center cropped to
the final image size. 1.14 was taken from the widely adopted practice of resizing to 256 before taking a
square crop of 224. This is common in ImageNet [
The final image sizes used are as follows: - BEiT-Base/p16: 3842 - DINOv2-Base: 4342 - DINOv2-Large: 5182 - SAM-ViT-Huge: 10242
A two-layer network was trained to project the concatenated embeddings into a new embedding that
better discriminated between the classes. Using the labels for each augmented image per instance, the
network was trained using PyTorch to project the concatenated embeddings into an embedding with
dimensionality of 768. The model consists of an input layer mapped to a hidden layer with dimensionality
2048, followed by an output layer with dimensionality 768. Both layers are fully connected, with ReLU
activation after the first layer. A batch size of 64 was used. The AdamW optimizer [
Multiple embedding pipelines are combined into an ensemble for the final predictions. For each embedding pipeline in the ensemble, the softmax of the cosine similarities between the prototype embedding for each class and the test embedding were calculated. The softmax probabilities per embedding pipeline in the ensemble were then averaged to get the final class probabilities. The mean inference wall-clock time for 5 seeds was 7.83 milliseconds per observation. For an ensemble, this scales linearly with the number of pipelines unless inference is performed in parallel.
The solution described in this study achieved 1st place on the private leaderboard for FungiCLEF 2025. This section details the results of ablations for the various components of the solution described in the previous section. For ablation experiments, models were trained using a split of the oficial training set into new training and validation subsets (used for early stopping), and evaluated on the oficial validation set (treated as a test set). Unless explicitly stated otherwise (e.g., Table 8 showing the private leaderboard performance for the top teams), all results are reported on this oficial validation set. The baseline for each of these ablations is a single embedding pipeline (instead of the final ensemble) with the same seed for the training-validation split and projection network initialization. All ablation experiments use deterministic seeding as described in Section 4.4.
As shown in Table 2 and Table 3, the inclusion of train and test time augmentations are the largest contributors to the performance of this solution. The inclusion of train time augmentations without test time augmentations results in a Top-5 accuracy reduction of 26.4 percentage points while the removal of both train time and test time augmentations results in a reduction of 27.1 percentage points.
Learning a projection of the concatenated embeddings improves model performance as shown in Table 4. The projection networks utilized by our top-ranking solution were trained with a combination of cross-entropy and infoNCE losses. Table 5 shows that this combined loss outperforms either loss alone.
Both combining models at the feature level and also ensembling predictions from multiple learned projections of those embeddings improve performance. Table 6 shows that concatenating the embeddings from multiple pretrained models outperforms using the embedding from a single pretrained model. DINOv2-Large proves to be a particularly strong performer as a single model. Conversely, SAM-ViT-H performs quite poorly without the context of the other embedding models. It appears that SAM-ViT-H can be removed from the embedding model combination to decrease the computational demands of the solution without degrading performance.
Table 7 shows that an ensemble of embedding pipelines outperforms a single embedding pipeline. As previously described, each member of the ensemble difered only by the training-validation split used to train the projection model and the random initialization of the projection model. For this ensemble, the seed for the training-validation split and the projection network initialization were diferent for each member of the ensemble, since otherwise predictions from the ensemble would be identical to that of a single pipeline.
To ensure reproducibility and statistical robustness, all ablation experiments used deterministic seeding. Each configuration was run with 5 independent replicates, and we report mean ± standard deviation.
Seeds were computed hierarchically as , = 1000 ⋅ + , (6) where ∈ {0, 1, 2, 3, 4} indexes the experimental replicate and ∈ {0, 1, … , −1} indexes the ensemble member ( = 0 for single pipelines, and is the ensemble size). This structure ensures non-overlapping seeds across replicates and ensemble members while maintaining reproducibility.
For single pipelines, the same seed was used for both the training-validation split and the projection model initialization. In ensembles, each member difered only by its corresponding seed, ensuring diversity through variation in both data splits and projection model initializations.
Private leaderboard performance for the top 10 ranking teams is shown in Table 8. Our models achieved the best performance for the competition metric (Top-5 accuracy).
Simple methods are suficient to achieve state-of-the-art performance for few-shot classification of fungi from image data. In this study, we described our winning approach for the FungiCLEF 2025 challenge. Using pretrained image classification and feature extraction networks, embeddings can be cached and subsequently used to train lightweight projection networks. These networks can be ensembled to further boost performance. Concatenation of embeddings from multiple frozen embedding models and averaging embeddings from multiple image augmentations perform well despite their simplicity. Importantly, we show that test-time augmentation is critical to the performance of this method.
During the preparation of this work, the author used Anthropic Claude Sonnet 4 in order to: Paraphrase and reword. After using this tool/service, the author reviewed and edited the content as needed and takes full responsibility for the publication’s content. [16] A. van den Oord, Y. Li, O. Vinyals, Representation learning with contrastive predictive coding, 2019. URL: https://arxiv.org/abs/1807.03748. arXiv:1807.03748. [17] K. Musgrave, S. J. Belongie, S.-N. Lim, Pytorch metric learning, ArXiv abs/2008.09164 (2020). [18] A. Kendall, Y. Gal, R. Cipolla, Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, 2018. URL: https://arxiv.org/abs/1705.07115. arXiv:1705.07115.