PaperlayerView exact revision on arXiv

Source-linked AI summary

CIA-SSD: Confident IoU-Aware Single-Stage Object Detector From Point Cloud

Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, Chi-Wing Fu

arXiv:2012.03015v1cs.CV

TL;DR

Single-stage point-cloud detectors can misalign localization accuracy and classification confidence. CIA-SSD addresses this with spatial-semantic feature aggregation, IoU-aware confidence rectification, and distance-variant IoU-weighted NMS, achieving state-of-the-art Moderate AP on KITTI among single-stage detectors.

  • Problem

    Single-stage point-cloud detectors can misalign localization accuracy and classification confidence, while robust predictions require both low-level spatial and high-level semantic features.

  • Method

    CIA-SSD adaptively fuses spatial and semantic features, rectifies classification confidence using predicted IoU, and applies distance-variant IoU-weighted NMS.

  • Results

    Moderate AP 80.28% on the KITTI test set, with top performance among existing single-stage detectors.

  • Takeaways & Limitations

    The detector combines more accurate predictions, confidence scores better aligned with localization, smoother regressions, and fewer redundant false positives.

  • Takeaways & Limitations

    KITTI validation and test results are inconsistent, and hard AP is lower than state-of-the-art test-set methods, possibly because the splits have mismatched distributions.

Abstract

from arXiv · show

Existing single-stage detectors for locating objects in point clouds often treat object localization and category classification as separate tasks, so the localization accuracy and classification confidence may not well align. To address this issue, we present a new single-stage detector named the Confident IoU-Aware Single-Stage object Detector (CIA-SSD). First, we design the lightweight Spatial-Semantic Feature Aggregation module to adaptively fuse high-level abstract semantic features and low-level spatial features for accurate predictions of bounding boxes and classification confidence. Also, the predicted confidence is further rectified with our designed IoU-aware confidence rectification module to make the confidence more consistent with the localization accuracy. Based on the rectified confidence, we further formulate the Distance-variant IoU-weighted NMS to obtain smoother regressions and avoid redundant predictions. We experiment CIA-SSD on 3D car detection in the KITTI test set and show that it attains top performance in terms of the official ranking metric (moderate AP 80.28%) and above 32 FPS inference speed, outperforming all prior single-stage detectors. The code is available at https://github.com/Vegeta2020/CIA-SSD.

1 Introduction

CIA-SSD targets the misalignment between localization accuracy and classification confidence in single-stage point-cloud detection. It combines spatial-semantic feature aggregation, IoU-aware confidence rectification, and distance-variant IoU-weighted NMS for accurate and efficient predictions.

  • Motivation: Single-stage detectors offer greater efficiency than two-stage detectors, motivating their use for time-sensitive point-cloud detection.Two-stage detectors generally achieve higher average precision through refinement, while single-stage detectors use simpler network structures.
  • Problem: Localization accuracy and classification confidence may misalign because existing detectors treat object localization and category classification as separate tasks.Two-stage methods address this issue by predicting IoUs for regressed boxes and refining confidence with soft IoU labels.
  • Problem: Single-stage detectors cannot effectively rectify confidence like two-stage networks because they lack second-stage features from predicted boxes and rely mainly on anchors or classified positive points.Their resulting IoU predictions may therefore be less accurate than those based on two-stage region proposals.
  • Method: CIA-SSD uses a convex IoU-aware confidence rectification function to increase discrimination between precise and imprecise box regressions.The method leverages anchor-feature-based IoU predictions to rectify confidence during post-processing.
  • Method: The lightweight SSFA module fuses high-level semantic and low-level spatial features, while DI-NMS uses depth-aware IoU weighting to smooth regressions and reduce redundant predictions.These components target more accurate box and confidence predictions and address distance-related redundant false positives.
  • Results: 80.28% moderate AP and 30.76 ms inference time are reported for CIA-SSD on 3D car detection in the KITTI test set.The paper presents these results as top performance and real-time speed among state-of-the-art single-stage detectors.

2 Related Work

Related work distinguishes two-stage detectors, which refine region proposals, from simpler single-stage detectors that directly regress scores and boxes. The paper positions CIA-SSD as an efficient single-stage detector motivated by competitive precision and real-time potential.

  • Detector Categories: Two-stage detectors generate region proposals and refine them in a second-stage network, whereas single-stage detectors directly regress class scores and bounding boxes.The distinction centers on refinement and network simplicity.
  • Two-Stage Detectors: PointRCNN, Part-A2, and STD represent two-stage approaches using proposal generation and region-level refinement strategies.The supplied passage identifies PointRCNN's anchor-free proposals, Part-A2's sparse convolutional backbone and RoI-aware pooling, and STD's spherical anchors with a segmentation branch.
  • Qualitative Comparison: CIA-SSD is illustrated as producing better-aligned boxes and fewer redundant predictions than SASSD in an example.The comparison is presented as a qualitative illustration of detection output quality.
  • Single-Stage Detectors: VoxelNet, PointPillar, and SECOND represent single-stage approaches using voxel, pillar, or sparse-convolution-based point-cloud feature extraction.The supplied passage describes VoxelNet's voxel feature encoding and PointPillar's pillar-based 2D convolutional detection architecture.
  • Motivation: Recent single-stage detectors are described as achieving comparable average precision to state-of-the-art two-stage detectors while retaining high efficiency.This combination motivates developing CIA-SSD for top performance and real-time speed relative to previous single-stage detectors.

3 Confident IoU-Aware Single Stage Detector

CIA-SSD combines spatial-semantic feature aggregation, IoU-aware confidence rectification, and distance-variant IoU-weighted NMS in a single-stage point-cloud detector. Its pipeline uses sparse point-cloud encoding, multi-task prediction, confidence adjustment, and post-processing to improve localization and suppress redundant predictions.

  • Overview: CIA-SSD comprises SPConvNet encoding, SSFA feature extraction, a confidence-rectifying multi-task head, and DI-NMS post-processing.The multi-task head performs classification and localization before confidence rectification and final suppression.
  • Spatial-Semantic Feature Aggregation: SSFA adaptively fuses high-level semantic and low-level spatial features for bounding-box and classification-confidence prediction.Its spatial and semantic branches are combined through learned BEV attention maps and element-wise addition.
  • IoU-Aware Confidence Rectification: Anchor-based IoU predictions remain discriminative despite imperfect accuracy, with high predicted IoUs often associated with high real IoUs.This association supports distinguishing precise from imprecise bounding-box regressions.
  • IoU-Aware Confidence Rectification: The rectification item g = i^β suppresses low-IoU predictions and enhances high-IoU predictions before modifying the classification score.The paper reports more discriminative high-IoU predictions with β = 4 and defines the confidence function as f = c · i^β.
  • Distance-Variant IoU-Weighted NMS: DI-NMS refines confidence using anchor-to-prediction BEV distance and applies IoU-weighted suppression during post-processing.The method targets distant-object oscillations and redundant false-positive predictions caused by sparse point observations.

4 Experiments

CIA-SSD is evaluated for KITTI car detection against state-of-the-art detectors, with ablations measuring its data processing, SSFA, confidence, and DI-NMS components. It achieves strong test-set accuracy and fast inference, while validation results show a distribution-related limitation.

  • 4.1 Implementation Details: The KITTI evaluation uses 7,481 training samples, 7,518 test samples, car detection, and a 0.7 IoU threshold for AP.The training samples are divided into 3,712 training and 3,769 validation samples.
  • 4.2 Comparison with State-of-the-Arts: CIA-SSD ranks first among single-stage detectors with 80.28% moderate AP and 89.59% easy AP on the KITTI test set.It outperforms recent single-stage detectors by about 0.5–0.8 points in moderate AP and most recent two-stage detectors by about 0.6–1.8 points.
  • 4.2 Comparison with State-of-the-Arts: Validation performance is slightly below some state-of-the-art results for easy and moderate AP, and hard AP is below test-set state-of-the-art methods.The authors attribute this inconsistency to mismatched KITTI validation and test distributions.
  • 4.3 Ablation Study: Global and local augmentation improve easy and moderate AP, while ground-truth augmentation, similar-object training, and difficulty filtering improve all AP levels.These techniques establish a strong baseline for validating the proposed modules.
  • 4.3 Ablation Study: SSFA improves moderate, easy, and hard AP by 0.37, 0.44, and 0.32, respectively, while increasing GPU occupation by about 10%.The corresponding SASSD module increases GPU occupation by about 27%, supporting SSFA’s lightweight design.
  • 4.3 Ablation Study: The confidence function improves easy, moderate, and hard AP by 0.20, 0.46, and 0.76, with β = 4 producing the highest PCC and moderate AP.Without score rectification, AP still increases from 79.17 to 79.30 because IoU prediction makes features aware of relative box-ground-truth locations.
  • 4.3 Ablation Study: DI-NMS raises easy, moderate, and hard AP by 0.38, 0.18, and 0.16, respectively, but its gain is limited by imperfect 3D object coverage.CIA-SSD has a 30.76 ms average inference time, including 24.33 ms in-network processing, and is faster than the compared single-stage and two-stage detectors.

5 Conclusion

CIA-SSD combines spatial-semantic feature aggregation, confidence rectification, and distance-variant IoU-weighted NMS for point-cloud detection. It achieves state-of-the-art KITTI Moderate AP among single-stage detectors while maintaining real-time efficiency.

  • CIA-SSD achieves state-of-the-art 3D detection performance on KITTI Moderate AP compared with existing single-stage detectors.
  • CIA-SSD uses spatial-semantic feature aggregation to extract robust features for object predictions.
  • Its confidence function rectifies classification scores to alleviate misalignment between localization accuracy and classification confidence.
  • Distance-variant IoU-weighted NMS produces smoother results and avoids redundant zero-IoU false positives.
  • CIA-SSD attains real-time detection efficiency and runs fastest among the compared recent state-of-the-art detectors.
Loading 2012.03015v1…