[Paper Review] CIA-SSD: Confident IoU-Aware Single-Stage Object Detector From Point Cloud

Zheng Wu, Weiliang Tang+3 more|arXiv (Cornell University)|Dec 5, 2020

Advanced Neural Network Applications33 references37 citations

TL;DR

CIA-SSD introduces a confident IoU-aware single-stage detector for point clouds, using spatial-semantic feature fusion, IoU-based confidence rectification, and distance-variant NMS to achieve state-of-the-art KITTI moderate AP with real-time speed.

ABSTRACT

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.

Motivation & Objective

• Motivate efficient and accurate 3D object detection in LiDAR point clouds with a single-stage detector.
• Address misalignment between localization precision and classification confidence in single-stage detectors.
• Propose lightweight modules to fuse spatial and semantic features and rectify confidence without a second stage.
• Introduce DI-NMS to smooth regressions and reduce redundant predictions.
• Demonstrate state-of-the-art performance on KITTI and real-time inference on a single GPU.

Proposed method

• Spatial-Semantic Feature Aggregation (SSFA) to fuse high-level semantic and low-level spatial features.
• IoU-Aware Confidence Rectification to align classification confidence with localization quality without a second stage.
• Distance-variant IoU-weighted NMS (DI-NMS) to smooth bounding-box regressions and suppress redundant predictions.
• Anchor-based IoU prediction branch trained with IoU targets, used to rectify confidence during testing via a Confidence Function.
• Loss formulation incorporating classification, box regression, direction, and IoU prediction losses.
• Sparse convolutional backbone (SPConvNet) guided by SECOND for BEV feature extraction.

Experimental results

Research questions

• RQ1 Can a lightweight SSFA module improve localization and classification stability in a single-stage 3D detector?
• RQ2 Does IoU-aware confidence rectification improve the alignment between localization accuracy and detection confidence in a single-stage framework?
• RQ3 Can distance-aware NMS further reduce false positives and stabilize regressions in 3D detection from point clouds?
• RQ4 How does CIA-SSD perform on KITTI against both single-stage and two-stage detectors in terms of moderate AP and speed?

Key findings

• CIA-SSD achieves 89.59% Easy and 80.28% Moderate AP on the KITTI test set (LiDAR, 3D car detection).
• CIA-SSD achieves 72.87% Hard AP on KITTI test set for 3D car detection.
• On KITTI validation, CIA-SSD attains 90.04 Easy, 79.81 Moderate, and 78.80 Hard AP, indicating strong generalization.
• SSFA contributes notable gains by enhancing feature fidelity with modest overhead (~10-27% GPU usage compared to baseline).
• IoU-aware confidence rectification with beta=4 improves Moderate AP and increases Pearson correlation between IoU and confidence.
• DI-NMS yields improvements in Easy/Moderate/Hard AP (e.g., Easy +0.38, Moderate +0.18, Hard +0.16) and provides smoother post-processing.
• CIA-SSD runs at 30.76 ms average inference time, outperforming other contemporary single-stage detectors and all listed two-stage baselines.