AAI_2025_Capstone_Chronicles_Combined

timestamps, and locations where the images were captured. For this project, we focus on a subset of ~65,000 images with bounding box annotations, which are essential for object detection and model training (Leorna & Brinkman, 2022). To train and evaluate our models, we cropped out the animal objects from the original images using the bounding box coordinates and resized them to a uniform shape of 224x224 pixels. These cropped images are then used for both a custom CNN model and a fine-tuned ResNet-18 model. The dataset is split into two stages for model development. First, images are categorized by location, using images from 100 locations as in-distribution dataset and the remaining 40 locations as out-of-distribution dataset. Then, a time-based split is applied, with images from the first year used for training and validating the baseline models. The remaining years are reserved to simulate production. This approach simulates real-world deployment, where the model must generalize to data captured at future times and locations.

282