Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Selective Focus: Investigating Semantics Sensitivity in Post-training Quantization for Lane Detection

Yunqian Fan, Xiuying Wei, Ruihao Gong, Yuqing Ma, Xiangguo Zhang, Qi Zhang, Xianglong Liu

TL;DR

This work tackles the gap between post-training quantization and the semantic-rich post-processing used in lane detection. It introduces the Lane Distortion Score to quantify how quantization-induced errors distort lanes and reveals intra-head and inter-head semantic sensitivities that standard PTQ ignores. The authors propose the Selective Focus framework, consisting of Semantic Guided Focus and Sensitivity Aware Selection, to inject post-processing information into PTQ by foreground-focused reconstruction and dynamic head selection. Across diverse LD models and datasets, the approach achieves notable F1 gains (up to 6.4% under 4-bit activation) with significant speedups, demonstrating practical impact for edge deployment of lane detection systems.

Abstract

Lane detection (LD) plays a crucial role in enhancing the L2+ capabilities of autonomous driving, capturing widespread attention. The Post-Processing Quantization (PTQ) could facilitate the practical application of LD models, enabling fast speeds and limited memories without labeled data. However, prior PTQ methods do not consider the complex LD outputs that contain physical semantics, such as offsets, locations, etc., and thus cannot be directly applied to LD models. In this paper, we pioneeringly investigate semantic sensitivity to post-processing for lane detection with a novel Lane Distortion Score. Moreover, we identify two main factors impacting the LD performance after quantization, namely intra-head sensitivity and inter-head sensitivity, where a small quantization error in specific semantics can cause significant lane distortion. Thus, we propose a Selective Focus framework deployed with Semantic Guided Focus and Sensitivity Aware Selection modules, to incorporate post-processing information into PTQ reconstruction. Based on the observed intra-head sensitivity, Semantic Guided Focus is introduced to prioritize foreground-related semantics using a practical proxy. For inter-head sensitivity, we present Sensitivity Aware Selection, efficiently recognizing influential prediction heads and refining the optimization objectives at runtime. Extensive experiments have been done on a wide variety of models including keypoint-, anchor-, curve-, and segmentation-based ones. Our method produces quantized models in minutes on a single GPU and can achieve 6.4% F1 Score improvement on the CULane dataset.

Selective Focus: Investigating Semantics Sensitivity in Post-training Quantization for Lane Detection

TL;DR

This work tackles the gap between post-training quantization and the semantic-rich post-processing used in lane detection. It introduces the Lane Distortion Score to quantify how quantization-induced errors distort lanes and reveals intra-head and inter-head semantic sensitivities that standard PTQ ignores. The authors propose the Selective Focus framework, consisting of Semantic Guided Focus and Sensitivity Aware Selection, to inject post-processing information into PTQ by foreground-focused reconstruction and dynamic head selection. Across diverse LD models and datasets, the approach achieves notable F1 gains (up to 6.4% under 4-bit activation) with significant speedups, demonstrating practical impact for edge deployment of lane detection systems.

Abstract

Lane detection (LD) plays a crucial role in enhancing the L2+ capabilities of autonomous driving, capturing widespread attention. The Post-Processing Quantization (PTQ) could facilitate the practical application of LD models, enabling fast speeds and limited memories without labeled data. However, prior PTQ methods do not consider the complex LD outputs that contain physical semantics, such as offsets, locations, etc., and thus cannot be directly applied to LD models. In this paper, we pioneeringly investigate semantic sensitivity to post-processing for lane detection with a novel Lane Distortion Score. Moreover, we identify two main factors impacting the LD performance after quantization, namely intra-head sensitivity and inter-head sensitivity, where a small quantization error in specific semantics can cause significant lane distortion. Thus, we propose a Selective Focus framework deployed with Semantic Guided Focus and Sensitivity Aware Selection modules, to incorporate post-processing information into PTQ reconstruction. Based on the observed intra-head sensitivity, Semantic Guided Focus is introduced to prioritize foreground-related semantics using a practical proxy. For inter-head sensitivity, we present Sensitivity Aware Selection, efficiently recognizing influential prediction heads and refining the optimization objectives at runtime. Extensive experiments have been done on a wide variety of models including keypoint-, anchor-, curve-, and segmentation-based ones. Our method produces quantized models in minutes on a single GPU and can achieve 6.4% F1 Score improvement on the CULane dataset.
Paper Structure (42 sections, 2 theorems, 12 equations, 6 figures, 7 tables, 1 algorithm)

This paper contains 42 sections, 2 theorems, 12 equations, 6 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Lane Detection Models
  4. Post-traning Quantization
  5. Preliminaries
  6. Notation
  7. Problem Definition
  8. Method
  9. Semantic Sensitivity
  10. Sensitivity to post-process
  11. Lane Distortion Score
  12. Semantic Guided Focus
  13. Intra-head Semantic Sensitivity
  14. Intra-head Sensitivity Focus
  15. Sensitivity Aware Selection
  16. ...and 27 more sections

Key Result

Theorem 1

Given $\mathcal{M}$ representing a matrix function that discerns elements associated with foreground or background regions, and $\mathcal{C}$ denoting the confidence function of FP models, offering confidence scores for semantics linked with the foreground, the following inequation stands:

Figures (6)

  • Figure 1: The framework of Selective Focus. Three modules are designed to mine the semantics sensitivity in the post-training quantized lane detection. Semantic Sensitivity Measuring measures the semantic sensitivity quantitatively; Sensitivity Aware Selection adapts the optimization objectives according to dynamic sensitivity. Semantic Guided Focus enables PTQ to focus on the foreground with a practical proxy.
  • Figure 2: Example of confidence and semantics in post-process. (a) The confidence is used to predict whether a keypoint is at a certain location. (b) The offset is regressed to the shift between the real keypoint and its downscaled grid. Offset is only valid in the indices with positive confidence prediction.
  • Figure 3: Common types of lane distortion caused by slight perturbation and the proposed Lane Distortion Score. (a) Bend: unexpected shifts at the terminals; (b) Spike: shifts in the middle; (c) Misalignment: missing or extra lane points. Any kind of lane distortion can be represented as a combination of those three types of distortion. (d) We measure the distortion between lanes with two types of point relationships: match and mismatch.
  • Figure 4: Illustration of semantic sensitivity. The intra-head semantic sensitivity shows that the foreground is sensitive to perturbation while the background is not. The inter-head semantic sensitivity indicates that heads are of different importance for post-processing. More illustrations are listed in the Appendix A.
  • Figure 5: Illustration of inter-head semantic sensitivity of more lane detection models with more datapoints.
  • ...and 1 more figures

Theorems & Definitions (3)

  • Theorem 1
  • Theorem 2
  • proof