Quantization Robustness to Input Degradations for Object Detection
Toghrul Karimov, Hassan Imani, Allan Kazakov
2025-09-01
Summary
This paper investigates how shrinking the size of object detection models, like YOLO, to make them faster affects their ability to accurately identify objects when images aren't perfect, like when they're noisy or blurry.
What's the problem?
Making object detection models smaller and faster, through a process called post-training quantization, often means reducing the precision of the numbers the model uses. While this speeds things up, it can make the model less reliable when dealing with real-world images that have imperfections like noise, blur, or compression from being saved as a JPEG. The core issue is understanding how much these imperfections impact the accuracy of these smaller, faster models.
What's the solution?
The researchers tested different versions of YOLO models, ranging in size, using various levels of precision. They then exposed the models to images with seven different types of common image degradations – things like noise, blur, and JPEG compression – and a combination of all of them. They also tried a new method of 'training' the model for the smaller size by showing it both clean and degraded images during the calibration process, hoping this would make it more robust. They measured how well the models performed on clean and degraded images to see if the new calibration method helped.
Why it matters?
This work is important because it shows that simply making a model smaller doesn't automatically mean it will work well in the real world. While reducing precision does make models faster, it can significantly reduce accuracy when images aren't perfect. The research suggests that simply showing the model degraded images during calibration doesn't consistently fix this problem, and highlights the need for better techniques to ensure these smaller models remain reliable in everyday situations.
Abstract
Post-training quantization (PTQ) is crucial for deploying efficient object detection models, like YOLO, on resource-constrained devices. However, the impact of reduced precision on model robustness to real-world input degradations such as noise, blur, and compression artifacts is a significant concern. This paper presents a comprehensive empirical study evaluating the robustness of YOLO models (nano to extra-large scales) across multiple precision formats: FP32, FP16 (TensorRT), Dynamic UINT8 (ONNX), and Static INT8 (TensorRT). We introduce and evaluate a degradation-aware calibration strategy for Static INT8 PTQ, where the TensorRT calibration process is exposed to a mix of clean and synthetically degraded images. Models were benchmarked on the COCO dataset under seven distinct degradation conditions (including various types and levels of noise, blur, low contrast, and JPEG compression) and a mixed-degradation scenario. Results indicate that while Static INT8 TensorRT engines offer substantial speedups (~1.5-3.3x) with a moderate accuracy drop (~3-7% mAP50-95) on clean data, the proposed degradation-aware calibration did not yield consistent, broad improvements in robustness over standard clean-data calibration across most models and degradations. A notable exception was observed for larger model scales under specific noise conditions, suggesting model capacity may influence the efficacy of this calibration approach. These findings highlight the challenges in enhancing PTQ robustness and provide insights for deploying quantized detectors in uncontrolled environments. All code and evaluation tables are available at https://github.com/AllanK24/QRID.