Network quantization allows inference to be conducted using low-precision arithmetic for improved inference efficiency of deep neural networks on edge devices. However, designing aggressively low-bit (e.g., 2-bit) quantization schemes on complex tasks, such as object detection, still remains challenging in terms of severe performance degradation and unverifiable efficiency on common hardware. In this paper, we propose an Accurate Quantized object Detection solution, termed AQD, to fully get rid of floating-point computation. To this end, we target using fixed-point operations in all kinds of layers, including the convolutional layers, normalization layers, and skip connections, allowing the inference to be executed using integer-only arithmetic. To demonstrate the improved latency-vs-accuracy tradeoff, we apply the proposed methods on RetinaNet and FCOS. In particular, experimental results on MS-COCO dataset show that our AQD achieves comparable or even better performance compared with the full-precision counterpart under extremely low-bit schemes, which is of great practical value.

中文翻译：

---

AQD：迈向精确量化目标检测

网络量化允许使用低精度算法进行推理，以提高边缘设备上深度神经网络的推理效率。然而，在复杂任务（例如对象检测）上设计积极的低位（例如，2 位）量化方案在严重的性能下降和常见硬件上无法验证的效率方面仍然具有挑战性。在本文中，我们提出了一种准确的量化对象检测解决方案，称为 AQD，以完全摆脱浮点计算。为此，我们的目标是在所有类型的层中使用定点运算，包括卷积层、归一化层和跳过连接，允许仅使用整数算法执行推理。为了演示改进的延迟与准确度权衡，我们将所提出的方法应用于 RetinaNet 和 FCOS。特别是在 MS-COCO 数据集上的实验结果表明，我们的 AQD 在极低位方案下实现了与全精度对应物相当甚至更好的性能，具有很大的实用价值。