SqueezeSegV2: Improved Model Structure and Unsupervised Domain Adaptation for Road-Object Segmentation from a LiDAR Point Cloud

Abstract

Earlier work demonstrates the promise of deeplearning-based approaches for point cloud segmentation; however, these approaches need to be improved to be practically useful. To this end, we introduce a new model SqueezeSegV2 that is more robust to dropout noise in LiDAR point clouds. With improved model structure, training loss, batch normalization and additional input channel, SqueezeSegV2 achieves significant accuracy improvement when trained on real data. Training models for point cloud segmentation requires large amounts of labeled point-cloud data, which is expensive to obtain. To sidestep the cost of collection and annotation, simulators such as GTA-V can be used to create unlimited amounts of labeled, synthetic data. However, due to domain shift, models trained on synthetic data often do not generalize well to the real world. We address this problem with a domainadaptation training pipeline consisting of three major components: 1) learned intensity rendering, 2) geodesic correlation alignment, and 3) progressive domain calibration. When trained on real data, our new model exhibits segmentation accuracy improvements of 6.0-8.6% over the original SqueezeSeg. When training our new model on synthetic data using the proposed domain adaptation pipeline, we nearly double test accuracy on real-world data, from 29.0% to 57.4%. Our source code and synthetic dataset will be open-sourced.

PASS3D: Precise and Accelerated Semantic Segmentation for 3D Point Cloud

“…As shown in Tab.I, our method outperforms previous state-of-the-art methods with remarkable margins in 'Pedestrian' and 'Cyclist' categories, raising 16.5% and 17.2% respectively. We develop the average IoU from 44.9% of SqueezeSegv2 [6] to 52.8%. Our data augmentation method strikingly improves the performance of all classes.…”

“…Existing works on 3D object detection or semantic segmentation based on point cloud data can be divided into three ways: 1) 2D-based methods: Inspired by mature image-based semantic segmentation frameworks, several methods project the point cloud into the BEV (birds-eye-view) ( [10], [3], [4], [18]) or FV (front-view) ( [5], [6], [7]) and use a 2D CNN to learn the characteristics of the point cloud for detection or semantic segmentation. In [3], a fast single-stage detector is designed, utilizing a specific-height-encoded BEV input.…”

“…These solutions are efficient enough, but without providing semantic information. Inspired by image-based semantic segmentation methods, some researchers [3], [4], [5], [6], [7] employ mature CNN-based neural network to predict the semantic label for each pixel by projecting 3D point clouds into 2D plane. Such methods like SqueezeSeg [5] and SqueezeSegv2 [6] are real-time but not optimal because of ignoring the internal geometric information in the 3D point cloud leading to insufficient performance.…”

LDLS: 3-D Object Segmentation Through Label Diffusion From 2-D Images

“…2) Comparison to Other Lidar Segmentation Methods: We benchmark LDLS through a comparison against Squeeze-Seg [6], SqueezeSegV2 [7], and PointSeg [8], state-of-the-art convolutional neural network methods for object segmentation in lidar point clouds. These methods take as input a lidar point cloud transformed through panoramic projection into a 64 × 512 × 5 tensor, where the 5 channels are x-, y-, and zcoordinates, depth, and lidar intensity.…”

LU-Net: An Efficient Network for 3D LiDAR Point Cloud Semantic Segmentation Based on End-to-End-Learned 3D Features and U-Net

“…In [19], [20] and [21], the authors use a 5-channel rangeimage as input of their network. These 5 channels are made of the 3D coordinates (x, y, z), the reflectance (r) and the spherical depth (d).…”