Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mixed Precision PointPillars for Efficient 3D Object Detection with TensorRT

Ninnart Fuengfusin, Keisuke Yoneda, Naoki Suganuma

TL;DR

Real-time LIDAR 3D object detection on edge devices is challenged by wide numerical distributions and outliers. The authors introduce a mixed-precision framework for PointPillars that uses PTQ to identify sensitive layers via layer-wise INT8 quantization and AP evaluation, then greedily assigns FP16 to the top-$k$ sensitive layers and finalizes with PTQ or QAT calibration, all compatible with TensorRT. PTQ alone can achieve competitive performance without training, while QAT closes the gap to FP32; deployment on TensorRT yields latency reductions up to $2.35\times$ and size reductions up to $2.26\times$, with only about $2.018\%$ of parameters being in sensitive layers. A key insight is that very small calibration sets (e.g., six frames) reduce outlier effects in PTQ, improving accuracy. Overall, the method enables efficient, high-performance 3D detection on edge hardware, demonstrated on KITTI with strong practical implications for autonomous systems.

Abstract

LIDAR 3D object detection is one of the important tasks for autonomous vehicles. Ensuring that this task operates in real-time is crucial. Toward this, model quantization can be used to accelerate the runtime. However, directly applying model quantization often leads to performance degradation due to LIDAR's wide numerical distributions and extreme outliers. To address the wide numerical distribution, we proposed a mixed precision framework designed for PointPillars. Our framework first searches for sensitive layers with post-training quantization (PTQ) by quantizing one layer at a time to 8-bit integer (INT8) and evaluating each model for average precision (AP). The top-k most sensitive layers are assigned as floating point (FP). Combinations of these layers are greedily searched to produce candidate mixed precision models, which are finalized with either PTQ or quantization-aware training (QAT). Furthermore, to handle outliers, we observe that using a very small number of calibration data reduces the likelihood of encountering outliers, thereby improving PTQ performance. Our methods provides mixed precision models without training in the PTQ pipeline, while our QAT pipeline achieves the performance competitive to FP models. With TensorRT deployment, our models offer less latency and sizes by up to 2.35 and 2.26 times, respectively.

Mixed Precision PointPillars for Efficient 3D Object Detection with TensorRT

TL;DR

Real-time LIDAR 3D object detection on edge devices is challenged by wide numerical distributions and outliers. The authors introduce a mixed-precision framework for PointPillars that uses PTQ to identify sensitive layers via layer-wise INT8 quantization and AP evaluation, then greedily assigns FP16 to the top- sensitive layers and finalizes with PTQ or QAT calibration, all compatible with TensorRT. PTQ alone can achieve competitive performance without training, while QAT closes the gap to FP32; deployment on TensorRT yields latency reductions up to and size reductions up to , with only about of parameters being in sensitive layers. A key insight is that very small calibration sets (e.g., six frames) reduce outlier effects in PTQ, improving accuracy. Overall, the method enables efficient, high-performance 3D detection on edge hardware, demonstrated on KITTI with strong practical implications for autonomous systems.

Abstract

LIDAR 3D object detection is one of the important tasks for autonomous vehicles. Ensuring that this task operates in real-time is crucial. Toward this, model quantization can be used to accelerate the runtime. However, directly applying model quantization often leads to performance degradation due to LIDAR's wide numerical distributions and extreme outliers. To address the wide numerical distribution, we proposed a mixed precision framework designed for PointPillars. Our framework first searches for sensitive layers with post-training quantization (PTQ) by quantizing one layer at a time to 8-bit integer (INT8) and evaluating each model for average precision (AP). The top-k most sensitive layers are assigned as floating point (FP). Combinations of these layers are greedily searched to produce candidate mixed precision models, which are finalized with either PTQ or quantization-aware training (QAT). Furthermore, to handle outliers, we observe that using a very small number of calibration data reduces the likelihood of encountering outliers, thereby improving PTQ performance. Our methods provides mixed precision models without training in the PTQ pipeline, while our QAT pipeline achieves the performance competitive to FP models. With TensorRT deployment, our models offer less latency and sizes by up to 2.35 and 2.26 times, respectively.
Paper Structure (16 sections, 4 equations, 3 figures, 5 tables)

This paper contains 16 sections, 4 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Proposed Method
  4. Prerequisites
  5. PointPillars
  6. INT8 Quantization and TensorRT
  7. Mixed Precision PointPillars
  8. PTQ Sensitive Layer Search
  9. INT8 and Latency
  10. Data Calibration and Performance
  11. Experimental Results and Discussion
  12. Sensitive Layer Search
  13. Mixed Precision with PTQ and QAT
  14. Runtime Latency
  15. Model Size
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Overview of PointPillars architectures and how to decide the datatype of a layer based on its sensitivity.
  • Figure 2: Left: The relationship between the mean of 3D AP40 across three classes with moderate difficulty (mAP40) and the number of data calibration samples. Right: The relationship between the maximum observed input value in the first convolutional layer of PointPillars and the number of data calibration samples. Note that the x-axis is provided in log scale.
  • Figure 3: Sensitive layer search. Y-axis indicates mAP40. x-axis displays the layer index to assign as INT8 layer, where all other layers are FP.