Object detection models show mixed robustness to quantization and input degradations

A new study investigates how post-training quantization (PTQ) affects the robustness of YOLO object detection models when faced with real-world input degradations like noise and blur. Researchers evaluated various precision formats, including Static INT8, and proposed a degradation-aware calibration strategy. While Static INT8 offers significant speedups, the proposed calibration method did not consistently improve robustness across most models and degradations, though some benefits were seen in larger models under specific noise conditions. AI

影响 Provides insights into deploying quantized object detection models in uncontrolled environments, highlighting challenges in robustness.

排序理由 Academic paper evaluating model robustness to input degradations.

在 arXiv cs.CV 阅读 →

• arXiv
• COCO
• FP16
• ONNX
• FP32
• TensorRT
• YOLO
• INT8
• UINT8

• 论文
• 其他

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →