Efficient Fine Tuning for Fashion Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Fashion Technology

Background:

• Pre-trained models excel in object detection but struggle with domain-specific data and dataset noise.
• Specialized fields like fashion imaging require robust zero-shot capabilities, which are often limited by existing datasets.
• Challenges include missed and false detections in clothing object detection tasks.

Purpose of the Study:

• To address limitations in zero-shot object detection for fashion imaging.
• To introduce a novel, large-scale clothing object detection benchmark, Garment40K.
• To propose an efficient fine-tuning method to improve clothing target detection.

Main Methods:

• Construction of the Garment40K dataset: over 140,000 human images and 40,000 clothing images with detailed annotations.
• Development of an efficient fine-tuning approach based on the Grounding DINO framework.
• Integration of additional similarity loss constraints and adapter modules into the Grounding DINO model, creating Improved Grounding DINO.

Main Results:

• The Garment40K dataset provides a rich resource with 2 major categories and 15 fine-grained subclasses of clothing.
• Improved Grounding DINO significantly enhances the detection of clothing targets, reducing missed and false detections.
• Fine-tuning adapter modules requires minimal computational cost while achieving performance comparable to full parameter fine-tuning.

Conclusions:

• The Garment40K dataset and Improved Grounding DINO model offer substantial advancements for computer vision in the clothing domain.
• The proposed method enables efficient deployment on low-cost visual sensors, broadening accessibility.
• This work improves the accuracy and efficiency of object detection in specialized fashion imaging applications.