Novel View Synthesis for Few-Shot Object Detection

Sammanfattning

This thesis study the use of different Neural Radiance Field (NeRF) methods, specifically Instant-NGP, as a data-centric augmentation for Few-Shot Object Detection (FSOD). Motivated by surveillance scenarios such as UAV reconnaissance, where annotated views of novel objects are sparse, the work investigates whether synthetic views generated from limited images can improve object detection performance. A simulation-based pipeline was developed to train NeRF-methods under controlled conditions and project annotations onto rendered views. These synthetic images were used to fine-tune a Faster R-CNN detector trained on base classes. Results show that NeRF-augmented fine-tuning improves average precision on novel classes, provided the azimuthal separation between input views is under 18°. Beyond this threshold, image quality and detector performance degrade due to rendering artifacts and loss of structural consistency. Training solely on synthetic views without fine-tuning proved ineffective. The findings demonstrate both the potential and limitations of NeRF-based augmentation in FSOD.