AAI_2025_Capstone_Chronicles_Combined

annual update cycles can achieve stable cross-geographic performance, providing a practical framework for multi-location deployment while highlighting the need for domain-specific calibration to address the residual geographic performance gap. Conclusion WildScan delivers a fully functional wildlife detection and classification pipeline—developed entirely without transfer learning and driven by our custom ScratchResNet architecture—integrated with MegaDetector v5 for robust object localization under low-light, occluded, and cluttered conditions. We have demonstrated this capability through a Gradio-based application that guides users through a streamlined workflow: uploaded images are first processed by MegaDetector v5 to detect and localize animal subjects, which are then automatically cropped to isolate each individual. The ScratchResNet model classifies these crops, providing top species predictions accompanied by confidence scores. Results are displayed with overlaid bounding boxes and labels alongside a ranked list of predictions, and a built-in feedback mechanism allows users to flag incorrect detections or classifications. All user annotations are captured and fed back into the training pipeline, enabling WildScan to continuously refine its performance and adapt to new data in real time.

297