SalsaNext: Fast, Uncertainty-Aware Semantic Segmentation of LiDAR Point Clouds

By Tiago Cortinhal, George Tzelepis, Eren Erdal Aksoy. International Symposium on Visual Computing, 2020

557 citations recorded by the source corpus.

In this paper, we introduce SalsaNext for the uncertainty-aware semantic segmentation of a full 3D LiDAR point cloud in real-time. SalsaNext is the next version of SalsaNet [1] which has an encoder-decoder architecture where the encoder unit has a set of ResNet blocks and the decoder part combines upsampled features from the residual blocks. In contrast to SalsaNet, we introduce a new context module, replace the ResNet encoder blocks with a new residual dilated convolution stack with gradually increasing receptive fields and add the pixel-shuffle layer in the decoder. Additionally, we switch from stride convolution to average pooling and also apply central dropout treatment. To directly optimize the Jaccard index, we further combine the weighted cross-entropy loss with Lovasz-Softmax loss [2]. We finally inject a Bayesian treatment to compute the epistemic and aleatoric uncertainties for each point in the cloud. We provide a thorough quantitative evaluation on the Semantic-KITTI dataset [3], which demonstrates that the proposed SalsaNext outperforms other state-of-the-art semantic segmentation networks and ranks first on the Semantic-KITTI leaderboard. We also release our source code this https URL.

Subfields: Computational Mechanics, Computer Vision and Pattern Recognition, Environmental Engineering

View Paper: https://www.semanticscholar.org/paper/7662ea3aa5c78921c0c2032e81d3a4fd7ce1c3ba

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Mechanics

1. Xinge Zhu
2. Dahua Lin
3. Tai Wang
4. Yushi Zhu
5. Jie Sun
6. Jianyun Xu
7. Ruixiang Zhang
8. Shiliang Pu

In Computer Vision and Pattern Recognition

1. Hongsheng Li
2. Dahua Lin
3. Xinge Zhu
4. Tai Wang
5. Xiaogang Wang
6. Shiliang Pu
7. Luc Van Gool
8. Dacheng Tao

In Environmental Engineering

1. Dahua Lin
2. Tai Wang
3. Xinge Zhu
4. Hongsheng Li

Citations in the corpus

• Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution (2020, unknown): “[3] 50.4 376.3 102 (78+24) 49.9 SqueezeSegV3-21 [68] 9.4 187.5 97 (73+24) 51.6 SqueezeSegV3-53 [68] 26.2 515.2 238 (214+24) 55.9 3D-MiniNet [1] 4.0 { { 55.8 PolarNet [76] 13.6 135.0 62 57.2 SalsaNext [10] 6.7 62.8 71 (47+24) 59.5 SPVNAS (Ours) 1.1 8.9 89 60.3 Table 3. Results of outdoor scene segmen”
• RPVNet: A Deep and Efficient Range-Point-Voxel Fusion Network for LiDAR Point Cloud Segmentation (2021, methodology): “SalsaNext [12] takes Salsanet [1] as baseline and presents an uncertainty-aware mechanism for point feature learning.”
• Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation (2020, methodology): “Most existing outdoor-scene point cloud segmentation methods [16, 10, 27, 1, 48, 20] focus on converting the 3D point cloud to 2D grids, to enable the usage of 2D Convolutional Neural Networks.”