Voxel R-CNN: Towards High Performance Voxel-based 3D Object Detection

Deng, Jiajun; Shi, Shaoshuai; Li, Peiwei; Zhou, Wengang; Zhang, Yanyong; Li, Houqiang

doi:10.48550/arxiv.2012.15712

Cited by 33 publications

(72 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…cos(δ), are zero, which could stop backpropagation learning at the extreme opposite output. To resolve this problem, researchers [50,51,52,53,54,55,56,57,58,59,60,61] performed regression on both the cosine and sine of the angle (cos(θ), sin(θ)). When converting any given 1D Scalar-Based Orientation θ into 2-Dimensional values, they can be mapped to 2-D Cartesian system with angle's cosine value projected to x-axis and sine value projected to y-axis.…”

Section: Alpha (Local/allocentric Rotation)mentioning

confidence: 99%

SoK: Vehicle Orientation Representations for Deep Rotation Estimation

Tu¹,

Peng²,

Leung³

et al. 2021

Preprint

View full text Add to dashboard Cite

In recent years, there is an influx of deep learning models for 3D vehicle object detection. However, little attention was paid to orientation prediction. Existing research work proposed various vehicle orientation representation methods for deep learning, however a holistic, systematic review has not been conducted. Through our experiments, we categorize and compare the accuracy performance of various existing orientation representations using the KITTI 3D object detection dataset [1], and propose a new form of orientation representation: Tricosine. Among these, the 2D Cartesian-based representation [cos(θ), sin(θ)], or Single Bin, achieves the highest accuracy, with additional channeled inputs (positional encoding and depth map) not boosting prediction performance. Our code is published on GitHub: https://github.com/umd-fire-coml/KITTI-orientationlearning

show abstract

Section: Alpha (Local/allocentric Rotation)mentioning

confidence: 99%

SoK: Vehicle Orientation Representations for Deep Rotation Estimation

Tu¹,

Peng²,

Leung³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…These representations are then fed into 2D CNN architectures to regress bounding box locations and categories. Another popular lidar-based approach explores various representations of points using 3D convolutional networks such as voxels [11,61,65,74], spheres [32,46], pillars [25], or learning point features directly [35,47] using point operators [39,40]. These features are transformed to a BEV feature map where 3D boxes are generated using sparse convolution or transformer architectures [33,35].…”

Section: Introductionmentioning

confidence: 99%

See Eye to Eye: A Lidar-Agnostic 3D Detection Framework for Unsupervised Multi-Target Domain Adaptation

Tsai,

Berrio,

Shan

et al. 2021

Preprint

View full text Add to dashboard Cite

Sampling discrepancies between different manufacturers and models of lidar sensors result in inconsistent representations of objects. This leads to performance degradation when 3D detectors trained for one lidar are tested on other types of lidars. Remarkable progress in lidar manufacturing has brought about advances in mechanical, solid-state, and recently, adjustable scan pattern lidars. For the latter, existing works often require fine-tuning the model each time scan patterns are adjusted, which is infeasible. We explicitly deal with the sampling discrepancy by proposing a novel unsupervised multi-target domain adaptation framework, SEE, for transferring the performance of state-of-the-art 3D detectors across both fixed and flexible scan pattern lidars without requiring fine-tuning of models by end-users. Our approach interpolates the underlying geometry and normalizes the scan pattern of objects from different lidars before passing them to the detection network. We demonstrate the effectiveness of SEE on public datasets, achieving state-ofthe-art results, and additionally provide quantitative results on a novel high-resolution lidar to prove the industry applications of our framework. This dataset and our code will be made publicly available.

show abstract

“…This is similar to the processing method that has been widely studied for 2D images, but the quantization loss of position information will inevitably occur in the voxelization process. These are called voxel-based approaches [5,6,14,40,[45][46][47]. On the other hand, with the PointNet and PointNet++ proposed by Qi et al [25,26], some method directly learn features from raw point-cloud and predict 3D bounding boxes by foreground points and their features.…”

Section: Introductionmentioning

confidence: 99%

“…PV-RCNN [28] is a representative work among them, and its performance has reached the state of the arts, but due to the introduction of complex point operations, it cannot meet the real-time requirements (10 fps). However, Voxel-RCNN [6] proposed that the traditional voxel-based methods which extract features from bird's eye view (BEV) ignore the 3D information, which reduces the performance of the detector. Therefore, the voxel RoI pooling operation is proposed, and the performance of the cat is similar or even better than that of PV-RCNN on KITTI dataset [8].…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Structure Information is the Key: Self-Attention RoI Feature Extractor in 3D Object Detection

Zhang¹,

Zheng²,

et al. 2021

Preprint

View full text Add to dashboard Cite

Unlike 2D object detection where all RoI features come from grid pixels, the RoI feature extraction of 3D point cloud object detection is more diverse. In this paper, we first compare and analyze the differences in structure and performance between the two state-of-the-art models PV-RCNN and Voxel-RCNN. Then, we find that the performance gap between the two models does not come from point information, but structural information. The voxel features contain more structural information because they do quantization instead of downsampling to point cloud so that they can contain basically the complete information of the whole point cloud. The stronger structural information in voxel features makes the detector have higher performance in our experiments even if the voxel features don't have accurate location information. Then, we propose that structural information is the key to 3D object detection. Based on the above conclusion, we propose a Self-Attention RoI Feature Extractor (SARFE) to enhance structural information of the feature extracted from 3D proposals. SARFE is a plug-and-play module that can be easily used on existing 3D detectors. Our SARFE is evaluated on both KITTI dataset and Waymo Open dataset. With the newly introduced SARFE, we improve the performance of the state-ofthe-art 3D detectors by a large margin in cyclist on KITTI dataset while keeping real-time capability. The code will be released at https://github.com/Poley97/SARFE.

show abstract

Voxel R-CNN: Towards High Performance Voxel-based 3D Object Detection

Cited by 33 publications

References 20 publications

SoK: Vehicle Orientation Representations for Deep Rotation Estimation

SoK: Vehicle Orientation Representations for Deep Rotation Estimation

See Eye to Eye: A Lidar-Agnostic 3D Detection Framework for Unsupervised Multi-Target Domain Adaptation

Structure Information is the Key: Self-Attention RoI Feature Extractor in 3D Object Detection

Contact Info

Product

Resources

About