ChiralFusion: correcting distributional and perceptual asymmetries in multi-modal 3D object detection via density-position chirality
Year: 2026
Type: article
Abstract: Multi-modal 3D object detection faces significant challenges stemming from the intrinsic asymmetries of sensor data, specifically the irregular density of virtual points and the geometric misalignment caused by depth completion errors. Existing methods often treat these discrepancies as generic noise, lacking a unified theoretical model to quantify and correct such systematic biases. To address these issues, this paper proposes ChiralFusion , a novel chirality-aware fusion framework that theoretically reformulates these multi-modal discrepancies as Density Chirality and Position Chirality . Instead of relying on stochastic operations, we design a Density Chirality Discard (DCD) block that utilizes dynamic scoring to selectively preserve informative features in sparse regions while filtering redundant noise, thereby ensuring a uniform point distribution. Furthermore, we develop a Position Chirality Suppression (PCS) block that operates in the high-dimensional feature space to rectify geometric positioning deviations and learn robust feature representations invariant to local chiral noise. Extensive experiments validate the effectiveness of our method, which outperforms 24 recent state-of-the-art methods on the KITTI and nuScenes benchmarks, respectively. On the KITTI test set, ChiralFusion delivers a 3D mean Average Precision ( \(mAP_{3D}\) ) of 86.72% and a fast inference speed of 49 ms , achieving competitive accuracy and efficiency for autonomous driving perception.
Source: Complex & Intelligent Systems
Authors: Wenyu Ji, Yang Li, Jiabao Wang, Wei Bai, Zhuang Miao
Institution: PLA Army Engineering University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,490
Award: G472437798
OpenAlex ID: W7201835985