Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AQD: Towards Accurate Fully-Quantized Object Detection

Peng Chen, Jing Liu, Bohan Zhuang, Mingkui Tan, Chunhua Shen

TL;DR

This paper proposes an Accurate Quantized object Detection solution, termed AQD, to fully get rid of floating-point computation, and targets 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.

Abstract

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 trade-off, 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. Source code and models are available at: https://github.com/ziplab/QTool

AQD: Towards Accurate Fully-Quantized Object Detection

TL;DR

This paper proposes an Accurate Quantized object Detection solution, termed AQD, to fully get rid of floating-point computation, and targets 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.

Abstract

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 trade-off, 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. Source code and models are available at: https://github.com/ziplab/QTool

Paper Structure

This paper contains 16 sections, 9 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Proposed method
  4. Preliminary
  5. Integer-aware Quantized Representation
  6. Quantization function
  7. Floating-point Free Quantization
  8. Convolutional and Fully-connected Layers
  9. Normalization Layers
  10. Skip Connections
  11. Other Layers
  12. Discussions
  13. Experiments
  14. Comparison with State-of-the-art Methods
  15. Ablation Study
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Illustration of a typical block in the detection network. Double slash lines indicate the positions where quantization applies. Note that in a fully-quantized network, $x$ can be represented in the format of Eq. (\ref{['eq:represnet-linear']}), which is the output $\overline{y}$ of the preceding block.
  • Figure 2: Distribution of input activations at the batch normalization layer in the detection heads of a 2-bit ResNet-18 FCOS detector. Level-x denotes that the predictions are made on the x-th pyramid level. Different levels of features show different batch of statistics.
  • Figure 3: Illustration of the proposed quantized detection heads design. (a) Conventional detection frameworks with shared group normalization or batch normalization tian2019fcoslin2017focal; (b) The proposed multi-level BN that privatizes batch normalization layers for different heads.