Infrared Domain Adaptation with Zero-Shot Quantization

2024-9-03
Sevsay, Burak
Object detection models are gaining popularity in daily life and industry, increasing the demand for real-time computation of these models. Model compression is an essential technique to achieve faster inference and smaller footprints. Quantization is a widely used model compression technique, reducing bit-width and enhancing efficiency at the cost of quantization error. Methods like post-training quantization and quantization-aware training require training data to minimize this error. However, training data may be inaccessible due to privacy concerns in applications such as surveillance and autonomous driving. In such cases, zero-shot quantization becomes necessary, as it applies quantization without the need for training data. Additionally, infrared imagery, which is more resilient to illumination and weather conditions, is often a part of these applications. To the best of our knowledge, zero-shot quantization in the infrared domain has not been investigated before. This thesis adapts batch normalization statistics-based zero-shot quantization for the infrared domain. This method aims to generate synthetic data by utilizing batch normalization statistics. We thoroughly investigated the data generation process to achieve optimal results for YOLOv8 and RetinaNet. For infrared adaptation, we fine-tuned models that were pretrained on RGB images using infrared images. The evaluation is based on comparing zero-shot quantization results with those from both full-precision models and post-training quantization. Additionally, we examined the effect of model size on zero-shot quantization. Our results show that batch normalization statistics-based zero-shot quantization is more effective in the infrared domain and is an essential method when training data is unavailable.
Citation Formats
B. Sevsay, “Infrared Domain Adaptation with Zero-Shot Quantization,” M.S. - Master of Science, Middle East Technical University, 2024.