Research map / Detection and segmentation / CNN features
Selective Focus: Investigating Semantics Sensitivity in Post-training Quantization for Lane Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Investigates semantic sensitivity of post-training quantization (PTQ) for lane detection models and proposes Selective Focus, combining Semantic Guided Focus and Sensitivity Aware Selection, guided by a new Lane Distortion Score.
Problem
Standard PTQ reconstructs raw head outputs uniformly and ignores the post-processing of lane detection models, whose semantic outputs (offsets, locations, angles) differ in sensitivity, so quantization can cause severe lane distortion.
Approach
Introduce a Lane Distortion Score to measure post-processing-induced lane distortion, then use the model's own confidence outputs as a proxy foreground mask for reconstruction (Semantic Guided Focus) and dynamically select the top-k most sensitive heads during PTQ (Sensitivity Aware Selection).
Key results
- Under W4A4 on CULane, the method improves F1 by over 3% to 6.4% relative to prior PTQ baselines across lane detection models (e.g., 76.31 F1 for GANet Mid vs. 73.27 without Focus).
- Quantized models are produced in minutes on a single GPU, about 6x faster than block-wise PTQ (e.g., 33 min vs. 112 min for QDrop on CondLaneNet Small).
- On CondLaneNet Small W8A4, the method reaches 75.56 F1 versus 74.76 for QDrop and 76.92 for QAT (LSQ+, 1303 min).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Selective Focus instead of QDrop under W8A4 | Lane detection PTQ · CULane, CondLaneNet Small | F1 score: 74.76 -> 75.56 | Helped |
| Use Selective Focus instead of QDrop under W8A4 | Lane detection PTQ · CULane, GANet Small | F1 score: 75.77 -> 76.32 | Helped |
| Use Selective Focus instead of QDrop under W8A4 | Lane detection PTQ · CULane, LSTR Small | F1 score: 60.34 -> 63.14 | Helped |
| Use Selective Focus instead of QDrop under W8A4 | Lane detection PTQ · CULane, BézierLaneNet Small | F1 score: 64.48 -> 68.98 | Helped |
| Use Selective Focus instead of QDrop under W8A4 | Lane detection PTQ · CULane, LaneATT Small | F1 score: 66.58 -> 69.85 | Helped |
| Use Selective Focus instead of QDrop under W8A4 | Lane detection PTQ · CULane, SCNN Small | F1 score: 66.85 -> 69.61 | Helped |
| Use Selective Focus instead of QDrop under W8A4 | Lane detection PTQ · CULane, RESA Small | F1 score: 67.27 -> 69.46 | Helped |
| Add Semantic Guided Focus (ablation) | Selective Focus on GANet Small · CULane, W4A4 | F1 score: 73.27 -> 76.31 | Helped |
| Add Sensitivity Aware Selection (ablation) | Selective Focus on GANet Small · CULane, W4A4 | F1 score: 75.30 -> 76.31 | Helped |
| Add Semantic Guided Focus and Sensitivity Aware Selection (ablation) | Basic network-wise alignment on GANet Small · CULane, W4A4 | F1 score: 72.98 -> 76.31 | Helped |
| Set top-k from 1 to 2 (ablation) | Sensitivity Aware Selection on GANet Small · TuSimple | F1 score: 91.66 -> 96.46 | Helped |
| Set top-k from 2 to 3 (ablation) | Sensitivity Aware Selection on GANet Small · TuSimple | F1 score: 96.46 -> 96.51 | Helped |
Surprising findings
- Standard IoU-based metrics overlook local lane distortions, motivating the Lane Distortion Score.
- Anchor-based LaneATT collapses under some W4A4/W8A4 PTQ baselines (e.g., 0.00 to 0.64 F1 for several methods), while the proposed method keeps it at 37.59.
Limitations
- The Lane Distortion Score is tailored to distorted lanes and applies only to lanes with similar characteristics.
- The mask proxy relies on the assumption that the confidence output approximates a binomial mask; the theorem's covariance argument is informal.
- Experiments are limited to lane detection (plus a small COCO and CurveLanes generalization check).
Experimental setup
Datasets: CULane TuSimple MS COCO CurveLanes
Benchmarks: CULane TuSimple MS COCO CurveLanes
Metrics: F1 score accuracy bbox mAP runtime (minutes)
Baselines: OMSE ACIQ AdaRound BRECQ QDrop LSQ+
Builds on: AdaRound BRECQ QDrop OMSE ACIQ
Closest papers by meaning
- GSENet:Global Semantic Enhancement Network for Lane Detection (AAAI 2024 · similarity 0.88)
- Lane2Seq: Towards Unified Lane Detection via Sequence Generation (CVPR 2024 · similarity 0.88)
- Rethinking Lanes and Points in Complex Scenarios for Monocular 3D Lane Detection (CVPR 2025 · similarity 0.87)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.87)
- A Hybrid Global-Local Perception Network for Lane Detection (AAAI 2024 · similarity 0.86)
- TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.