4DenoiseNet: Adverse Weather Denoising From Adjacent Point Clouds

Abstract

Reliable point cloud data is essential for perception tasks e.g. in robotics and autonomous driving applications. Adverse weather causes a specific type of noise to light detection and ranging (LiDAR) sensor data, which degrades the quality of the point clouds significantly. To address this issue, this letter presents a novel point cloud adverse weather denoising deep learning algorithm (4DenoiseNet). Our algorithm takes advantage of the time dimension unlike deep learning adverse weather denoising methods in the literature. It performs about 10% better in terms of intersection over union metric compared to the previous work and is more computationally efficient. These results are achieved on our novel SnowyKITTI dataset, which has over 40000 adverse weather annotated point clouds. Moreover, strong qualitative results on the Canadian Adverse Driving Conditions dataset indicate good generalizability to domain shifts and to different sensor intrinsics.

An improved point cloud denoising method in adverse weather conditions based on PP-LiteSeg network

“…Trained approaches, particularly those based on neural networks, have made significant strides in semantic segmentation applied to PCD. These approaches are tailored for general segmentation tasks, and they can be trained to segment highly abstract shapes, including noise patterns caused by rain and fog ( Seppänen, Ojala & Tammi, 2022 ). Successful methods employ voxelization techniques, such as VoxNet, designed by Maturana & Scherer (2015) .…”

“…However, the model has more continuous branching structures, leading to a slower inference speed. Seppänen, Ojala & Tammi (2022) introduced 4DenoiseNet, which utilizes a new k-nearest neighbor search convolution on continuous point clouds, incorporating temporal-dimension information to improve denoising performance for snow. However, its limitation lies in the requirement for PCDs with temporal continuity.…”

“…As deep learning algorithms advance, trained networks outperform conventional methods in terms of both segmentation quality and real-time performance. Common point cloud semantic segmentation methods include point-based approaches ( Seppänen, Ojala & Tammi, 2022 ) and projection-based methods ( Heinzler et al, 2020 ). When a spherical projection is used to transform 3D data into 2D images, a matured 2D real-time lightweight semantic segmentation network can also be applied to point cloud segmentation.…”

Exploring metadiscourse in university lectures in Hong Kong

“…Despite achieving commendable results in the single domain, there is a significant performance drop when transitioning to new domains [42,50]. The limited generalization capability hinders their facilitation of real-world applications [49,52,89]. In reality, LiDAR datasets are marred by significant variances, encompassing variations in data patterns due to different * The first two authors contributed equally to this work.…”

“…Recent works [7,41,50,81,89,96,108] resort to unsupervised domain adaptation (UDA) for utilizing training data from both source and target domains to optimize one parameter set. Nevertheless, they either focus on only the sharing mapping between two domains (by ignoring dis-joint classes) or directly merge source domain labels to align with the target domain [42,115].…”