Object Detection Data Synthesis via Box-to-Image Generation Based on Diffusion Models.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Diffusion models show promise for augmenting object detection datasets.
• Current methods struggle with generating complex scenes, multi-class objects, and occlusions.

Purpose of the Study:

• To introduce ODGEN, a novel method for generating high-quality, bounding-box-conditioned images for object detection data synthesis.
• To improve controllability and quality in complex scene generation.

Main Methods:

• Fine-tuning pre-trained diffusion models on domain-specific datasets.
• Utilizing synthesized visual prompts with spatial constraints and textual descriptions for control.
• Developing a dataset synthesis pipeline for evaluation.

Main Results:

• ODGEN demonstrates robustness in complex and domain-specific scenes.
• Incorporating ODGEN-generated data improved object detection performance by up to 25.3% mAP@.50:.95.
• Achieved up to 5.6% mAP@.50:.95 advantage over existing methods on general domains.

Conclusions:

• ODGEN effectively facilitates data synthesis for object detection, particularly in challenging scenarios.
• The proposed method significantly boosts the performance of object detectors like YOLOv5 and YOLOv7.
• Stable Box Diffusion, a general model trained on large datasets, covers numerous object categories.