GoodQ Achieves State-of-the-Art Zero-Shot Quantization for Object Detectors

This paper introduces GoodQ, a novel pipeline for Zero-Shot Quantization-Aware Training (ZSQ-OD) specifically designed for object detection models. ZSQ-OD is crucial for deploying models on edge devices where access to original training data is restricted. Unlike previous methods that rely on noise optimization for synthesizing training sets, GoodQ leverages readily available generative models.The researchers identified three key challenges when integrating generative models into ZSQ-OD: handling dense information with multiple instances per image, managing class-wise distribution imbalances, and mitigating noise from pseudo-labels during QAT. GoodQ tackles these by implementing an Information-Dense Prompting strategy for multi-instance image generation, an Intrinsic Distribution-Aware Selection to align with pre-trained class distributions, and a Teacher-guided Adaptive Noise Reduction to reduce QAT noise.Empirical results demonstrate that GoodQ achieves state-of-the-art performance in low-bit ZSQ (W4A4) and even extends quantization to extreme bit-widths (W3A3). The extensive analysis provided also sheds light on the factors contributing to GoodQ's effectiveness, making it a significant advancement for efficient object detector deployment.