Infrared Domain Adaptation with Zero-Shot Quantization

Burak Sevsay,Erdem Akagündüz

2024-08-26

Abstract:Quantization is one of the most popular techniques for reducing computation time and shrinking model size. However, ensuring the accuracy of quantized models typically involves calibration using training data, which may be inaccessible due to privacy concerns. In such cases, zero-shot quantization, a technique that relies on pretrained models and statistical information without the need for specific training data, becomes valuable. Exploring zero-shot quantization in the infrared domain is important due to the prevalence of infrared imaging in sensitive fields like medical and security applications. In this work, we demonstrate how to apply zero-shot quantization to an object detection model retrained with thermal imagery. We use batch normalization statistics of the model to distill data for calibration. RGB image-trained models and thermal image-trained models are compared in the context of zero-shot quantization. Our investigation focuses on the contributions of mean and standard deviation statistics to zero-shot quantization performance. Additionally, we compare zero-shot quantization with post-training quantization on a thermal dataset. We demonstrated that zero-shot quantization successfully generates data that represents the training dataset for the quantization of object detection models. Our results indicate that our zero-shot quantization framework is effective in the absence of training data and is well-suited for the infrared domain.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper primarily focuses on addressing the issue of applying zero-shot quantization techniques in the infrared domain. Specifically, the researchers explored how to perform model quantization by generating synthetic data in the absence of training data, thereby reducing the accuracy loss during the quantization process. The core contribution of the paper lies in demonstrating how zero-shot quantization can be applied to object detection models on infrared images and validating the effectiveness of this method. The research team used RetinaNet as the benchmark model and utilized batch normalization statistics to generate synthetic data for quantization calibration. They compared the impact of using only mean statistics versus using both mean and standard deviation statistics on quantization performance and found that using only mean statistics could achieve better quantization results. Additionally, the study compared the performance differences between zero-shot quantization and traditional training data-based quantization methods (i.e., post-training quantization). The results showed that the mean Average Precision (mAP) of zero-shot quantization was comparable to post-training quantization, with smaller accuracy deviations. In summary, the paper proposes a new method for effectively performing model quantization in the absence of actual training data and validates its applicability and reliability in the infrared domain. This is particularly important for application areas involving privacy-sensitive data, such as medical and security surveillance.

Infrared Domain Adaptation with Zero-Shot Quantization

Zero-Shot Detection with Transferable Object Proposal Mechanism.

Zero-shot Adversarial Quantization

ZeroQ: A Novel Zero Shot Quantization Framework

Domain Adaptation Meets Zero-Shot Learning: an Annotation-Efficient Approach to Multi-Modality Medical Image Segmentation

IntraQ: Learning Synthetic Images with Intra-Class Heterogeneity for Zero-Shot Network Quantization

Zero-Shot Day-Night Domain Adaptation with a Physics Prior

A Generalized Zero-Shot Quantization of Deep Convolutional Neural Networks via Learned Weights Statistics

TexQ: Zero-shot Network Quantization with Texture Feature Distribution Calibration

Infrared colorization with cross-modality zero-shot learning

Is In-Domain Data Really Needed? A Pilot Study on Cross-Domain Calibration for Network Quantization

Zero-Shot Sharpness-Aware Quantization for Pre-trained Language Models

Generative Zero-shot Network Quantization

Fine-grained Data Distribution Alignment for Post-Training Quantization

Learning Accurate Low-bit Quantization towards Efficient Computational Imaging

Genie: Show Me the Data for Quantization

A Soft Calibration Technique for Infrared Imaging Based on Physical Model of Thermal Radiation and Variational Bayesian Inference from Small Dataset

Hybrid and non-uniform quantization methods using retro synthesis data for efficient inference

Zero-shot Domain Adaptation without Domain Semantic Descriptors

Long-Range Zero-Shot Generative Deep Network Quantization

Zero-Shot Deep Domain Adaptation