A guide to designing a bear face recognition system

In this blog post we’ll delve into the technical development of a bear face recognition system which is a critical component of the bear identification system developed in close collaboration with the BearID Project NGO.

Our research and software tool will provide a replicable technique and general approach that can be applied to other species beyond bears, which could aid conservation efforts worldwide.

– BearID Project

For a comprehensive understanding of this project, please click on the pipeline overview below:

Project Scope

While many species boast distinctive fur patterns for identification, brown bears lack consistent and unique markings. Furthermore, their weight can fluctuate significantly between seasons and throughout their lifetimes. Consequently, facial recognition emerges as a valuable alternative for individual identification.

In this article, our attention is directed towards the final phase of the bear identification system, specifically the recognition of bear faces.

The initial stage entails acquiring a mapping capable of embedding bear faces into a high-dimensional space, facilitating the clustering of individuals.

Embedding bear faces into a high dimensional space

Subsequently, we leverage this learned mapping to execute queries and retrieve the closest matching individuals.

Retrieving closest individuals from the learned mapping

Our collaboration with The BearID Project aims to significantly enhance their current model performance, which currently stands at accuracy@1 (top-1): 0.649 and accuracy@5 (top-5): 0.707.

Provided Dataset

The BearID Project has compiled a collection of bear images, showcasing their facial features, captured over recent years in forests across British Columbia and Brooks Falls.

After the development of the bear face segmentation system, as detailed in a prior blog post, we successfully generated approximately 4700 bear face images, representing a total of 132 individuals.

Generated bear faces by the segmentation model

Bursts of images

When encountering a bear, photographs or camera traps often capture multiple images of the same individual in very similar poses. Proper handling of these bursts of images during data splitting is crucial. Neglecting this step may lead to train/test data leakage, which can cause the model to inaccurately overreport its performance.

Gallery / Provided labels by The BearID Project - A burst of images with the same individual

Individual counts distribution

The curated dataset primarily consists of only a few image faces for most bears, posing a potential challenge for accurate identification. Conversely, some bears have hundreds of image faces, largely derived from bursts of images captured during encounters. Below, we present a distribution plot of individual counts to provide further insight into the data.

Individual counts distribution - How many image faces per individual?

re-IDentification

Definition

Re-identification, often abbreviated as re-ID, involves identifying and tracking individual animals across different camera traps or instances in wildlife habitats.

Objective

The primary goal is to accurately identify and track individual animals over time, allowing researchers to study behavior, population dynamics, migration patterns, and other ecological factors.

Challenges

Animal re-identification faces unique challenges compared to person or object re-identification:

• Variability in Appearance: Animals may exhibit variations in appearance due to factors such as changes in fur color, markings, or physical condition (e.g., injuries).
• Environmental Conditions: Camera traps are often deployed in outdoor environments where lighting conditions, weather, and vegetation can vary significantly, affecting image quality and visibility of animals.
• Species Variability: Different species may exhibit diverse morphologies and behaviors, requiring specialized models and approaches for accurate re-identification.

Common approach

Similar to other re-identification tasks, animal re-identification often involves the following steps:

• Feature Extraction: Deep learning-based models, such as convolutional neural networks (CNNs), are used to extract discriminative features from images captured by camera traps. These features encode unique characteristics of individual animals, such as fur patterns or facial markings. For brown bears - which lack unique fur and body markings - facial markings are key.
• Matching: Extracted features are then compared across different camera traps or instances to identify and track individual animals. Matching algorithms, such as nearest neighbor search or clustering, are commonly used to find similar feature representations corresponding to the same animal across different images.

Applications

Animal re-identification has numerous applications in wildlife conservation and ecological research:

• Population Monitoring: Researchers can use re-identification data to estimate population sizes, monitor trends, and assess the impact of environmental changes or human activities on wildlife populations.
• Behavioral Studies: Long-term tracking of individual animals allows for detailed studies of behavior, movement patterns, habitat use, and social interactions within animal populations.
• Conservation Planning: Re-identification data can inform conservation strategies by identifying key habitats, migration corridors, and areas of high wildlife activity for targeted conservation efforts.

Overall, animal re-identification plays a vital role in understanding and managing wildlife populations, facilitating conservation efforts, and supporting ecological research initiatives.

Closed Set, Open Set and Disjoint Set

In the context of re-ID, understanding the concepts of open sets, closed sets, and disjoint sets is critical.

Closed Set Identification

In a closed set scenario, the system is trained to recognize a predefined set of classes or identities. For example, in a closed set re-identification task, the system is trained to identify a specific set of animals from a gallery of known individuals. The key characteristic of a closed set approach is that the identities present in the testing or operational phase are limited to those seen during training. In other words, the system only recognizes identities that it has been explicitly trained on.

Open Set Identification

In contrast to closed set identification, open set identification deals with scenarios where the testing data might contain identities not seen during training. The system needs to be able to recognize known identities (closed set) while also detecting and handling unknown or novel identities. This means the system must have the capability to distinguish between familiar identities (in-distribution data) and unfamiliar ones (out-of-distribution data). Open set identification systems often incorporate techniques like anomaly detection or thresholding to identify instances that don’t belong to any known class.

Disjoint Set Identification

Disjoint set identification refers to situations where the identities in the training and testing datasets are completely separate. In other words, there is no overlap between the identities seen during training and those encountered during testing. This scenario is common in real-world applications where the population of identities is constantly changing, such as surveillance systems in crowded areas or public spaces.

Summary

In summary, closed set identification deals with recognizing a fixed set of known identities, open set identification extends this to handle unknown or novel identities, and disjoint set identification involves training and testing on completely separate sets of identities. Each approach has its own challenges and requirements, and the choice depends on the specific application and the nature of the data.

Data Modeling

Data Splits

We opted to create two distinct splits to assess the performance of the identification system in real-world scenarios:

• Open-set split: This split includes a portion of newly introduced identities in the testing phase, simulating encounters with previously unseen entities.
• Disjoint-set split: In this split, the training and testing datasets comprise entirely different identities, mimicking scenarios where the system encounters novel entities during deployment

To avoid data leakage in the open-set split, we implemented a careful splitting strategy based on both camera reference and date. This ensures that bursts of images captured by the same camera at the same time are consistently grouped into the same split (train, validation, or test).

Metric Learning

Overview

Metric learning is a machine learning paradigm focused on learning a distance metric or similarity function directly from data. Instead of relying on predefined distance measures, metric learning algorithms aim to discover a distance metric that optimally represents the underlying structure or relationships within the data. The goal is to ensure that similar instances are mapped closer together in the learned metric space, while dissimilar instances are pushed farther apart. Metric learning has applications in various domains, including image retrieval, face recognition, clustering, and classification, where accurately capturing the similarity or dissimilarity between data points is crucial for task performance.

Learning a metric space to embed bear faces

Losses

Loss functions play an important role in metric learning as they guide the optimization process to learn effective distance metrics or embedding spaces. Let’s explore the importance of some common losses in metric learning:

Contrastive Loss

The Contrastive Loss encourages similar instances to be closer together and dissimilar instances to be pushed apart in the embedding space. It achieves this by penalizing pairs of similar instances that are far apart and pairs of dissimilar instances that are close together.

By using a contrastive loss, the model learns to map instances of the same class (or similar instances) close to each other while maximizing the distance between instances of different classes (or dissimilar instances).

Triplet Margin Loss

A Triplet Margin Loss builds upon contrastive loss by considering triplets of anchor, positive, and negative examples. It ensures that the distance between the anchor and the positive example is smaller than the distance between the anchor and the negative example by at least a margin.

The Triplet margin loss explicitly enforces relative distance relationships between instances, which can lead to more discriminative embeddings. It helps in dealing with the issue of intra-class variance and inter-class separability.

Circle Loss

A Circle Loss defines a circular decision boundary in the embedding space, with each class represented by a circle. The radius of the circle is dynamically adjusted based on the intra-class variations, ensuring that samples from the same class are pulled together within the circle while maintaining a margin from samples of other classes.

A circle Loss offers several advantages over traditional softmax-based losses, including better handling of intra-class variations, robustness to noisy data, and improved generalization to unseen classes. It has been shown to achieve state-of-the-art performance in various metric learning tasks, particularly in scenarios with large intra-class variations or class imbalances.

ArcFace Loss

The ArcFace loss is particularly effective for face recognition tasks. It enhances the discriminative power of the learned embeddings by introducing a margin-based angular penalty.

The ArcFace loss operates in a hypersphere embedding space, where the angle between the feature vectors and the corresponding class-specific hypersphere centers is optimized. This ensures that intra-class variations are minimized while inter-class variations are maximized, leading to improved classification accuracy and better generalization.

Losses Summary

In summary, losses in metric learning are crucial for guiding the optimization process towards learning effective embedding spaces. They encourage the model to embed similar instances close together while pushing dissimilar instances apart. Each loss function has its advantages and is suitable for different applications and tasks, ultimately contributing to improved performance in tasks such as image retrieval, face recognition, and clustering.

Evaluation Metrics

Main metric - Accuracy@k

Accuracy at k is an evaluation metric used to measure the correctness of the top-k predictions made by a model. It assesses how many of the correct answers fall within the top-k ranked predictions. This metric is commonly used in classification tasks, recommendation systems, and information retrieval tasks.

Here’s how accuracy at k works:

1. Top-k Predictions: After making predictions for a set of inputs (e.g., class labels, recommendations, or search results), the model ranks these predictions based on their confidence scores or relevance scores.
2. Correctness Evaluation: Accuracy at k evaluates the correctness of the top-k predictions by considering whether the correct answer is among these top-k predictions.
3. Calculation: The accuracy at k is calculated by dividing the number of correct predictions within the top-k ranked results by k.
4. Interpretation: Accuracy at k provides insights into how well the model performs when considering only the top-ranked predictions. A higher accuracy at k indicates that a larger fraction of the correct answers are included within the top-k predictions, suggesting better performance.

Accuracy@k is particularly useful in scenarios where only the top-ranked predictions are considered, such as recommendation systems where users are only shown a limited number of recommendations or search engines where users typically only view the top search results. It helps assess the effectiveness of the model in providing relevant and accurate predictions within the top-k ranked results.

We monitor the accuracy at different values of k, including accuracy@1, accuracy@3, accuracy@5, and accuracy@10, both during the training process and during evaluation.

Model Topology

The model architecture we adopted consists of a common pretrained backbone, complemented by a compact embedder head, as illustrated below:

Model Topology - Backbone (trunk) and embedder head

Training

The training process is executed across two GPUs for over 100 epochs. Throughout training, evaluation metrics are continuously monitored, and we save the weights of the best-performing model. Training halts if there’s no improvement in model performance beyond a specified threshold, determined by the patience parameter.

We adopt different learning rates for training the backbone and the heads. This decision stems from the fact that the backbone is typically pretrained on large datasets for feature extraction, whereas the embedder is trained from scratch for the specific task. Typically, we employ a learning rate ratio of 10 to 100 between the backbone and the embedder to effectively balance the learning rates and ensure optimal training dynamics.

Metric Space Visualization - Embeddings

We visualize the metric spaces using UMAP, a versatile manifold learning and dimension reduction algorithm, and save the visualizations after every epoch. This enables us to track and evaluate the improvement of the embedder over time.

Embeddings over training time

Hard Negative Mining

Hard negative mining specifically targets the difficult or misclassified samples that contribute the most to the loss function, ensuring that the model focuses on learning from these challenging examples.

Here’s why hard negative mining is important in metric learning:

• Enhancing Discriminative Power: Hard mining helps to identify and prioritize the training samples that are most informative for improving the model’s ability to discriminate between classes or categories. By focusing on the “hard” examples, which are often misclassified or have ambiguous class boundaries, the model can learn more discriminative features.
• Effective Utilization of Training Data: In large-scale datasets, not all samples contribute equally to the learning process. Hard mining allows the model to efficiently utilize the training data by emphasizing the most informative samples, leading to faster convergence and better generalization performance.
• Addressing Class Imbalance: In many real-world applications, classes may be imbalanced, with fewer examples of certain classes compared to others. Hard mining helps to mitigate the effects of class imbalance by ensuring that the model learns equally from both easy and challenging examples, thus improving the overall performance on underrepresented classes.
• Improving Robustness to Outliers and Noise: Hard mining encourages the model to focus on learning from challenging samples, which can help improve its robustness to outliers, noisy data, and variations in the input space. By effectively handling difficult examples, the model becomes more resilient to noise and generalizes better to unseen data.

Overall, hard negative mining is critical in metric learning because it enables the model to concentrate on the most informative and challenging training samples, leading to more discriminative embeddings and improved performance animal re-identification.

Baseline

A baseline was quickly established using a pretrained ResNet18 as the backbone and a Circle Loss for 1 epoch.

Visualizing the Learned Metric Space: Clusters Yet to Emerge

The approach shows great promise, and selecting the appropriate hyperparameters was key in maximizing the model’s performance

Hyperparameter Search

After configuring experiment tracking, we conducted a random hyperparameter search across the following parameter space:

• backbones: ResNet18, ResNet50, Convnext_tiny, Convnext_large
• losses: tripletmargin, circle, arcface
• learning rates
• weight decay
• mining strategies: easy, semi hard, hard
• embedder’s depth
• optimizers: Adam, SGD, etc
• embedding size; 512, 1024, 2048
• data augmentation steps: rotation, color jitter, etc

We randomly sampled configurations from this parameter space and conducted training sessions for a few days on two GPUs.

Best Model

The winning combination comprises a convnext_large backbone paired with an arcface loss, employing a hard mining strategy, and trained using the Adam optimizer.

Visualizing the Learned Metric Space: Clusters have emerged

This robust model excels in identifying bears with exceptional accuracy, even in disjoint and open set scenarios, representing a significant advancement over the existing solution developed by BearID. Additionally, we opted to forego the face alignment stage, deeming it unnecessary due to the superior performance of our current approach.

Conclusion

In this guide, we’ve detailed the development of an open-source, state-of-the-art model tailored for animal re-identification, specifically targeting brown bear faces. The comprehensive system, crafted for deployment by The BearID Project to monitor bear populations in Canada, represents a significant step forward in wildlife conservation efforts. With its adaptable design, this technology stands poised to transcend species boundaries, offering potential applications across diverse ecosystems. The scalability and versatility of the model hold promise for advancing conservation endeavors on a global scale.