ID 2507: Video-Based Species Recognition in Camera Trap Footage from the Amazon

Topic Description:

Camera traps are widely used in wildlife monitoring, but automated species recognition remains challenging due to low image quality, occlusions, and morphological similarities between species. Most existing approaches rely on single-frame classification, disregarding valuable temporal cues in animal movement. This project aims to improve species recognition by using video-based methods that incorporate movement patterns and pose estimation. Additionally, we will explore the integration of ecological prior knowledge (species distributions) and possibly audio data to enhance classification robustness.

Data:

The study will use video data from multiple sites in the Amazon region of Brazil, provided by the Norwegian University of Life Sciences (NMBU). The dataset covers diverse habitats, varying lighting conditions (day vs. night), and includes a long-tailed class distribution where some species occur rarely. Annotations are partially available, requiring strategies to handle incomplete labels and class imbalance.

Tasks:

• Develop and implement a preprocessing pipeline for video-based species detection and sequence generation.
• Train and evaluate at least two video-based species recognition models (e.g., 3D CNNs, pose-based methods).
• Implement a Bayesian framework to integrate ecological prior knowledge and model uncertainty.
• Perform a comprehensive evaluation of classification performance across environmental conditions.

Requirements:

• Strong programming skills (Python, PyTorch).
• Experience with deep learning, particularly in computer vision (CNNs, action recognition, or pose estimation).
• Basic knowledge of Bayesian methods is a plus.
• Interest in wildlife conservation and ecological modelling.
• Ability to work independently and collaborate with an interdisciplinary research team.