Project DetailsVisual object counting in crowded scenes is a major challenge in computer vision. This project focuses on zero-shot (exemplar-free) object counting, where the system automatically counts the most frequent objects without human bounding boxes or text prompts. In this work, we analyzed the state-of-the-art baseline, SAVE (2025), which leverages a local object detector backbone (YOLOv8) to drive a Visual Embedding Module and a Self-Attention Matching Module. We identified that local detection-based backbones struggle to capture global scene contexts and suffer in noisy, cluttered environments or under extreme scale variations. To bridge this gap, we developed SaveDINO, a novel model that replaces the local detector with self-supervised foundation vision model DINOv3, adapted efficiently using LoRA and a custom adapter module. We conducted a thorough quantitative and qualitative evaluation using the standard FSC-147 benchmark and a custom-made Noisy Real-World dataset. The results reveal a clear semantic split: while SAVE excels at dense, rigid patterns, SaveDINO demonstrates superior robustness and steady performance when dealing with natural, deformable object categories. Finally, we deployed our models into an end-to-end interactive web demo system, making our findings inspectable and reproducible in real-time.
