We’re thrilled to launch AnimalCLEF25, a competition pushing the boundaries of AI-driven animal re-identification and biodiversity monitoring. Despite we provide a list of resources including WildlifeReID-10k, WildlifeDatasets, WildlifeTools, MegaDescriptor, and WildFusion, we want you to test some alternative approaches and methods to solve open-set recognition and other key challenges. Below we add some promising approaches to solving the open-set recognition problem in Animal-CLEF 2025.
\nChallenge: Standard classification models will struggle with unknown individuals.
\nChallenge: Many individuals have very few images, making it difficult for deep models to generalize.
\nChallenge: Re-identification models must be robust across different environments, camera settings, and species.
\nChallenge: Handling massive datasets with thousands of individuals efficiently.
\nWe encourage participants to explore these methods and share their insights. Are you working on a novel open-set recognition technique? Testing a new loss function? Let’s discuss innovative approaches to advancing the field!
\n🚀 What scientific methods are you planning to try? Share your thoughts below!
", "messages": [ { "id": "3151107", "postDate": "03/16/2025 09:59:50", "content": "We’re thrilled to launch AnimalCLEF25, a competition pushing the boundaries of AI-driven animal re-identification and biodiversity monitoring. Despite we provide a list of resources including WildlifeReID-10k, WildlifeDatasets, WildlifeTools, MegaDescriptor, and WildFusion, we want you to test some alternative approaches and methods to solve open-set recognition and other key challenges. Below we add some promising approaches to solving the open-set recognition problem in Animal-CLEF 2025.
\nChallenge: Standard classification models will struggle with unknown individuals.
\nChallenge: Many individuals have very few images, making it difficult for deep models to generalize.
\nChallenge: Re-identification models must be robust across different environments, camera settings, and species.
\nChallenge: Handling massive datasets with thousands of individuals efficiently.
\nWe encourage participants to explore these methods and share their insights. Are you working on a novel open-set recognition technique? Testing a new loss function? Let’s discuss innovative approaches to advancing the field!
\n🚀 What scientific methods are you planning to try? Share your thoughts below!
", "rawMarkdown": "We’re thrilled to launch **AnimalCLEF25**, a competition pushing the boundaries of **AI-driven animal re-identification and biodiversity monitoring**. Despite we provide a list of resources including [WildlifeReID-10k](https://arxiv.org/abs/2406.09211), [WildlifeDatasets](https://github.com/WildlifeDatasets/wildlife-datasets), [WildlifeTools](https://github.com/WildlifeDatasets/wildlife-tools), [MegaDescriptor](https://openaccess.thecvf.com/content/WACV2024/papers/Cermak_WildlifeDatasets_An_Open-Source_Toolkit_for_Animal_Re-Identification_WACV_2024_paper.pdf), and [WildFusion](https://arxiv.org/pdf/2408.12934), we want **you** to test some alternative approaches and methods to solve open-set recognition and other key challenges. Below we add some promising approaches to solving the open-set recognition problem in Animal-CLEF 2025.\n\n### **Open-Set Classification and Unknown Detection**\n**Challenge**: Standard classification models will struggle with unknown individuals.\n#### **Potential Approaches**:\n - **Open-Set Recognition Models**: Use techniques such as OpenMax, Outlier Exposure, or Energy-Based Models to handle unknown classes effectively.\n - **Uncertainty Estimation**: Implement Bayesian Neural Networks or Monte Carlo Dropout to measure prediction confidence and detect unknown individuals.\n - **Prototype-Based Methods**: Leverage class prototypes in embedding space and reject queries that exceed a threshold distance.\n - **Hybrid Architectures**: Combine CNNs with graph-based approaches (e.g., Graph Neural Networks) to model relationships between individuals.\n - **Clustering-Based Re-Identification**: Use self-supervised clustering techniques (e.g., DeepCluster, SwAV) to discover new identities.\n\n### **Few-Shot and Zero-Shot Learning**\n**Challenge**: Many individuals have very few images, making it difficult for deep models to generalize.\n#### **Potential Approaches**:\n - **Meta-Learning (e.g., MAML, Prototypical Networks)**: Train models to adapt quickly to unseen identities with minimal labeled data.\n - **Feature Transfer Learning**: Use pre-trained embeddings (e.g., CLIP, DINO) to generalize better across species and domains.\n - **Augmentation & Generative Methods**: Leverage GANs or diffusion models to synthesize additional examples for rare individuals.\n\n### **Domain Adaptation and Generalization**\n**Challenge**: Re-identification models must be robust across different environments, camera settings, and species.\n#### **Potential Approaches**:\n - **Adversarial Domain Adaptation**: Use methods like DANN (Domain Adversarial Neural Networks) or gradient reversal layers to learn domain-invariant features.\n - **Self-Supervised Representation Learning**: Train contrastive learning models (SimCLR, MoCo) on large-scale wildlife images for better generalization.\n - **Feature Normalization and Disentanglement**: Separate domain-specific and identity-specific features using disentanglement techniques.\n\n### **Scalable and Efficient Large-Scale Recognition**\n**Challenge**: Handling massive datasets with thousands of individuals efficiently.\n#### **Potential Approaches**:\n - **Efficient Indexing Methods**: Use Approximate Nearest Neighbors (FAISS, HNSW) for fast retrieval in large-scale embedding spaces.\n - **Hierarchical Recognition Models**: Implement a multi-stage approach that first classifies species and then identifies individuals within each species.\n - **Incremental Learning**: Use continual learning techniques (e.g., EWC, GEM) to adapt dynamically to new data without catastrophic forgetting.\n\n### **Discussion & Future Directions**\nWe encourage participants to explore these methods and share their insights. Are you working on a novel open-set recognition technique? Testing a new loss function? Let’s discuss innovative approaches to advancing the field!\n\n🚀 **What scientific methods are you planning to try?** Share your thoughts below!", "votes": null }, { "id": "3154206", "postDate": "03/19/2025 16:45:31", "content": "@picekl Thank you for the competition. I've been working on computer vision tasks for a while now, but never crossed paths with re-identification task. Looks a bit complex but very interesting to me. Personally I've decided to start from prototype-based techniques as I consider the provided pretrained model to be sufficient to improve the baseline score without altering the model itself, but implementing various post and preprocessing techniques. Now I am working on building a prototype for each particular individual and try comparing the resultant extracted features to the prototype vectors of known individuals and reject any vectors based on certain level of lack of similarity.
", "rawMarkdown": "picekl Thank you for the competition. I've been working on computer vision tasks for a while now, but never crossed paths with re-identification task. Looks a bit complex but very interesting to me. Personally I've decided to start from prototype-based techniques as I consider the provided pretrained model to be sufficient to improve the baseline score without altering the model itself, but implementing various post and preprocessing techniques. Now I am working on building a prototype for each particular individual and try comparing the resultant extracted features to the prototype vectors of known individuals and reject any vectors based on certain level of lack of similarity.", "votes": null }, { "id": "3154219", "postDate": "03/19/2025 16:56:09", "content": "Awesome @danielvalyano ! Thanks for sharing the details.
\nIn general, individual animal identification (i.e., animal re-identification) is quite similar to human facial recognition. There's a wealth of literature on this topic that explores various methods and techniques. You might find it useful to check out.
\nWe also plan to publish better baselines during the competition so participants can learn something hopefully useful, so stay tuned! :)
\nBest,
\nLukas
It´s an interesting comparison, would never think it´s somewhat similar, but now you´ve given me some hints for my tomorows´ submissions ;) I am also planning to team up with some contenders and post our own baselines on the go.
\nThanks again!
Hi there, I am new for this can you recommend relevant papers to read
", "rawMarkdown": "Hi there, I am new for this can you recommend relevant papers to read", "votes": null }, { "id": "3154628", "postDate": "03/20/2025 07:16:32", "content": "\nThere are a lot of \"proprietary\" approaches for single species re-id. Since there is a lot, for those, I recommend checking Google Scholar. For more general solutions, you can see some of our papers that focus on all in one species re-id.
\n\nBest,
\nLukas
@picekl Thank you for the competition. I've been working on computer vision tasks for a while now, but never crossed paths with re-identification task. Looks a bit complex but very interesting to me. Personally I've decided to start from prototype-based techniques as I consider the provided pretrained model to be sufficient to improve the baseline score without altering the model itself, but implementing various post and preprocessing techniques. Now I am working on building a prototype for each particular individual and try comparing the resultant extracted features to the prototype vectors of known individuals and reject any vectors based on certain level of lack of similarity.
", "votes": null, "replies": [ { "id": 3154219, "author_name": "picekl", "author_url": "", "post_date": "03/19/2025 16:56:09", "content": "Awesome @danielvalyano ! Thanks for sharing the details.
\nIn general, individual animal identification (i.e., animal re-identification) is quite similar to human facial recognition. There's a wealth of literature on this topic that explores various methods and techniques. You might find it useful to check out.
\nWe also plan to publish better baselines during the competition so participants can learn something hopefully useful, so stay tuned! :)
\nBest,
\nLukas
It´s an interesting comparison, would never think it´s somewhat similar, but now you´ve given me some hints for my tomorows´ submissions ;) I am also planning to team up with some contenders and post our own baselines on the go.
\nThanks again!
Hi there, I am new for this can you recommend relevant papers to read
", "votes": null, "replies": [ { "id": 3154628, "author_name": "picekl", "author_url": "", "post_date": "03/20/2025 07:16:32", "content": "\nThere are a lot of \"proprietary\" approaches for single species re-id. Since there is a lot, for those, I recommend checking Google Scholar. For more general solutions, you can see some of our papers that focus on all in one species re-id.
\n\nBest,
\nLukas