High-Quality Proposal Encoding and Cascade Denoising for Imaginary Supervised Object Detection

The introduction of Cascade HQP-DETR marks a significant advancement in the field of Imaginary Supervised Object Detection (ISOD), which has been challenged by the need for large-scale annotated datasets that are often expensive and labor-intensive to produce. Current methods suffer from limitations including simplistic prompts, poor image quality, and weak supervision. Cascade HQP-DETR aims to overcome these issues by employing a high-quality data pipeline that leverages LLaMA-3, Flux, and Grounding DINO to create the FluxVOC and FluxCOCO datasets, transitioning ISOD from weak to full supervision. This innovative approach enhances the training process by initializing object queries with image-specific priors, thereby accelerating convergence and promoting the learning of transferable features. The model was trained for just 12 epochs, indicating a potential for rapid deployment in real-world scenarios. By addressing the shortcomings of existing methods, this research could significant…