Technical Summary: GoodQ – Zero-Shot Quantization for Object Detectors using Off-the-Shelf Generative Models

1. Problem Statement

The deployment of Object Detection (OD) models on resource-constrained edge devices necessitates model compression via quantization. Standard Quantization-Aware Training (QAT) requires access to the original training dataset to fine-tune the model. However, in many real-world scenarios, access to original data is prohibited due to privacy regulations and security concerns.

While Zero-Shot Quantization (ZSQ) has emerged to address this by synthesizing proxy datasets, existing ZSQ methods for OD face significant limitations:

• Low-Bit Performance Degradation: Current ZSQ approaches for OD, such as TSOD (which relies on noise optimization), struggle to maintain performance in extreme low-bit regions (e.g., W4A4 and W3A3).
• Data Diversity: Noise-optimized synthetic datasets often lack the diversity required to approximate the original data distribution effectively, leading to excessive performance drops when quantization bit-widths are reduced.
• Task-Specific Challenges: OD differs fundamentally from image classification. It requires images with dense semantic information (multiple instances per image), handles inherently imbalanced class distributions, and relies on precise spatial localization (bounding boxes). Standard generative approaches often fail to address these specific constraints.

2. Methodology: The GoodQ Framework

The authors propose GoodQ (Generative off-the-shelf models for object detector Quantization), a pipeline that leverages off-the-shelf generative models (specifically diffusion models) tailored to the unique challenges of ZSQ-OD. The framework operates in three distinct stages:

Stage 1: Information-Dense Prompting (IDP)

To address the Information Density Challenge (the need for multiple objects per image), GoodQ moves beyond single-object prompts used in classification.

• Strategy: It employs a specific prompt template: "A {D} photo of {Q} {C1} and {Q} {C2}", where {D} is a descriptive adjective, {Q} is a quantitative adjective (e.g., "multiple"), and {C1}, {C2} are class labels.
• Outcome: This generates a large pool (160k images) of synthetic images containing multiple object instances and bounding boxes, better reflecting the density of real OD datasets like MS-COCO.

Stage 2: Intrinsic Distribution-Aware Selection (IDAS)

To address the Class-wise Imbalance Challenge (the skewed distribution of object classes in OD datasets), GoodQ selects a representative subset from the generated pool without access to original labels.

• Distribution Estimation: Since original data is unavailable, the method estimates the target class distribution by analyzing the bias terms of the pre-trained detection head. These bias parameters implicitly capture class priors.
• Selection Algorithm: A greedy selection algorithm iteratively chooses images to match the estimated target distribution. It prioritizes under-represented classes and selects images with the highest number of bounding boxes to ensure density, effectively curating a dataset that mirrors the long-tail distribution of the original OD data.

Stage 3: Teacher-guided Adaptive Noise Reduction (TANR)

To address the Pseudo-Label Challenge (noise introduced by assigning labels to synthetic images via a pre-trained detector), GoodQ modifies the QAT loss function.

• Soft Targets: Instead of relying on hard pseudo-labels (one-hot vectors) which introduce supervision noise, the method uses the soft predictions (logits) from a full-precision "Teacher" model as the ground truth targets.
• Adaptive Weighting: The loss function incorporates a QFocal-based Adaptive Weighting (AW) mechanism. This allows the student model to focus more on instances that are susceptible to quantization-induced noise, effectively mitigating the adverse effects of label noise during training.

3. Key Contributions

1. Identification of Challenges: The paper systematically identifies three specific hurdles when applying off-the-shelf generative models to ZSQ-OD: information density, class imbalance, and pseudo-label noise.
2. Novel Framework (GoodQ): It introduces a tailored pipeline comprising Information-Dense Prompting, Intrinsic Distribution-Aware Selection, and Teacher-guided Adaptive Noise Reduction to overcome these hurdles.
3. State-of-the-Art Performance: The framework achieves state-of-the-art results in low-bit ZSQ (W4A4) and extends successful quantization to extreme bit-widths (W3A3), outperforming optimization-based methods like TSOD.
4. Comprehensive Analysis: The authors provide extensive ablation studies and visualizations (e.g., UMAP of CLIP embeddings) demonstrating that data diversity and distribution alignment are critical factors for robust low-bit quantization.

4. Experimental Results

Experiments were conducted on YOLOv5 and YOLOv11 models using the MS-COCO 2017 validation set.

• Low-Bit Robustness: In the W4A4 regime, GoodQ significantly outperforms TSOD. For example, on YOLOv5-s, GoodQ achieves 21.0 mAP compared to TSOD's 18.4 mAP. In the extreme W3A3 regime, GoodQ maintains 10.2 mAP, whereas TSOD drops to 5.0 mAP.
• High-Bit Performance: GoodQ maintains competitive performance in higher bit-widths (W8A8, W6A6), achieving results comparable to methods that utilize real data subsets or noise-optimized datasets.
• Ablation Findings:
  • Diversity: Diffusion-generated datasets show higher feature diversity (measured by cosine distance) than noise-optimized datasets, correlating with better low-bit performance.
  • Distribution: The IDAS selection strategy significantly outperforms random selection or naive "many bounding box" selection, confirming that matching the intrinsic class distribution is crucial.
  • Noise Reduction: The TANR component consistently improves performance, particularly on diffusion-generated data where label noise is more prevalent.

5. Significance and Claims

The paper claims that GoodQ represents a significant advancement in Zero-Shot Quantization for Object Detection by successfully bridging the gap between the capabilities of off-the-shelf generative models and the specific requirements of detection tasks.

The authors emphasize that their approach allows for the deployment of highly compressed OD models (down to 3-bit) on edge devices without requiring access to the original training data. By addressing the specific structural and statistical challenges of OD (density, imbalance, and label noise), GoodQ overcomes the performance degradation that has historically plagued low-bit quantization in this domain. The work suggests that maximizing training data diversity and accurately aligning class distributions are the underlying factors contributing to the efficacy of generative-based ZSQ-OD.