1Dept. of Chemistry and Bioscience, Aalborg University · 2Dept. of Architecture, Design and Media Technology, Visual Analysis and Perception Lab, Aalborg University
IET Computer Vision, 20(1), e70091 (2026) · DOI: 10.1049/cvi2.70091 · Open access
Manual labelling of animal images remains a significant bottleneck in ecological research, limiting the scale and efficiency of biodiversity monitoring. This study investigates whether state-of-the-art Vision Transformer (ViT) foundation models can reduce thousands of unlabelled animal images directly to species-level clusters. We present a comprehensive benchmarking framework evaluating five ViT models combined with five dimensionality reduction techniques and four clustering algorithms across 60 species (30 mammals and 30 birds), with each test using a random subset of 200 validated images per species. Our primary, fully unsupervised finding is that DINOv3 embeddings with t-SNE and HDBSCAN reach a species-level V-measure of 0.943 with no prior knowledge of the number of species, rejecting only 1.14% of images as outliers. Assigning one label per cluster reduces manual labelling actions by 98.3%, with 94.6% of clustered images labelled correctly. When the true number of species is supplied, supervised hierarchical clustering raises this to 0.958, reported as an upper bound. We further demonstrate robustness to realistic long-tailed distributions and show that intentional over-clustering can reveal intra-specific variation including age classes, sexual dimorphism, and pelage differences.
ECCV 2026 Workshop on Computer Vision for Ecology (CV4Ecology)
A useful wildlife census is more than a count: population dynamics depend on sex ratios and age structure. Using red deer as a test case, we fuse RGB and thermal drone video at every stage with self-supervised DINOv3 features, confirming an animal only when both cameras agree and reading life stage from geo-referenced body size. Across four flights spanning the antler season, the fused pipeline correctly classifies 25 of 26 individuals, against 20 of 26 for either sensor alone, with multimodal species classification reaching 96.0%.
ECCV 2026 Workshop on Computer Vision for Ecology (CV4Ecology)
Fine-grained wildlife classification in aerial imagery is limited by unreliable labels as much as by model performance. We study adult-male identification in red deer, where the antler cycle defines predictable windows of reliable evidence. Using 7,295 RGB-only, thermal-only, and matched RGB+thermal crop sets labelled by three annotators, we show that seasonal structure links annotation quality, downstream classification, and selective prediction. Uncertainty-band abstention raises covered accuracy to 98.9%.
State-of-the-art models like SpeciesNet provide predictions across thousands of species but use conservative rollup strategies, leaving many animals labeled at high taxonomic levels. We present a hierarchical re-classification system combining SpeciesNet predictions with CLIP embeddings and metric learning to refine labels toward species-level identification, achieving 96.5% accuracy on re-classified detections.
Evaluates zero-shot approaches for organizing unlabeled camera trap imagery using self-supervised vision transformers (CLIP, DINOv2, MegaDescriptor) with clustering and continuous 1D similarity ordering. DINOv2 with UMAP and GMM achieves 88.6% accuracy, while 1D t-SNE sorting reaches 95.2% coherence for fish and 88.2% for mammals/birds. Deployed in production on Animal Detect to accelerate annotation workflows.
Works best on cameratrap images of mammals and birds
Supports JPG, PNG, WebP formats (max 5MB)
Wildlife Detection, MegaDetector, Camera Trap Analysis
TensorFlow, PyTorch, Wildlife Classification Models
Python, Wildlife APIs, LILA BC Datasets
Trail Cameras, IR Systems, Wildlife Monitoring
ROS, Hexapod Robots, Search & Rescue Systems
MSc Robotics, AI:EcoNet (PhD) - ongoing
Co-founded in 2023, Animal Detect is revolutionizing wildlife conservation through AI. Born from my passion for merging robotics with nature, this platform helps researchers process millions of camera trap images, turning wildlife data into actionable conservation insights.
Founded during my MSc at AAU with Eugene Galaxy, Really A Robot has been my 5+ year journey from robotics student to conservation technology entrepreneur. From hexapod search-rescue robots to wildlife cameras, it's where Animal Detect was born and my passion for technology-driven conservation flourished.