Pre-trained models have achieved success in object detection. However, challenges remain due to dataset noise and lack of domain-specific data, resulting in weaker zero-shot capabilities in specialized fields such as fashion imaging. We addressed this by constructing a novel clothing object detection benchmark, Garment40K, which includes more than 140,000 human images with bounding boxes and over 40,000 clothing images. Each clothing item within this dataset is accompanied by its corresponding category and textual description. The dataset covers 2 major categories, pants and tops, which are further divided into 15 fine-grained subclasses, providing a rich and high-quality clothing resource. Leveraging this dataset, we propose an efficient fine-tuning method based on the Grounding DINO framework to tackle the issue of missed and false detections of clothing targets. This method incorporates additional similarity loss constraints and adapter modules, leading to a significantly enhanced model named Improved Grounding DINO. By fine-tuning only a small number of additional adapter module parameters, we considerably reduced computational costs while achieving performance comparable to full parameter fine tuning. This allows our model to be conveniently deployed on a variety of low-cost visual sensors. Our Improved Grounding DINO demonstrates considerable performance improvements in computer vision applications in the clothing domain.
Keywords: clothing dataset; fine tuning; object detection.