EPGNet: Enhanced Point Cloud Generation for 3D Object Detection

Qingsheng Chen, Cien Fan, Weizheng Jin, Lian Zou, Fangyu Li, Xiaopeng Li, Hao Jiang, Minyuan Wu, Yifeng Liu2020 · DOI

TL;DR

This paper introduces EPGNet, a novel two-stage 3D object detection framework that leverages object symmetry to enhance point cloud representations. By generating "symmetry points," EPGNet effectively fills in missing information in sparse lidar scans, leading to improved accuracy in detecting objects, especially those that are occluded or distant.

What Problem Does This Paper Address?

3D object detection, particularly for autonomous driving, is crucial but challenging. Lidar sensors, while providing accurate depth information, often produce sparse point clouds due to scanning limitations and occlusions. This sparsity makes it difficult for existing methods to capture complete object structures and extract robust features, leading to performance degradation, especially for distant or partially hidden objects. While many methods exist, they often overlook inherent object properties like symmetry, which could be exploited to infer missing data and improve detection.

Methods

EPGNet employs a two-stage detection framework:

  1. Symmetry Point Generation Module:
    • Data Processing: The raw point cloud is voxelized and processed to extract non-empty voxel-wise features.
  2. Foreground Segmentation & Symmetry Point Prediction: A UNet-like encoder-decoder network is used to segment foreground points (potential objects) and predict the locations of their corresponding symmetry points. This is achieved by learning the spatial offsets of symmetry points relative to foreground points.
  3. Enhanced Point Cloud Creation: The original foreground points and the newly generated symmetry points are combined with non-empty voxel centers to form an "enhanced point cloud." This step aims to create a denser and more complete representation of objects.
  4. Region Proposal Network (RPN):
    • The enhanced point cloud is fed into a standard anchor-based RPN, similar to those used in 2D detection but adapted for 3D. This stage performs 3D box classification and regression to generate the final detection results.

The core novelty lies in the explicit prediction of symmetry points to augment the sparse point cloud data, leveraging the inherent geometric properties of objects like cars.

Datasets

  • KITTI Dataset: The primary dataset used for training and evaluation. The authors follow a common practice of splitting the KITTI training set into a training split (3712 samples) and a validation split (3769 samples) for experiments.

Key Insights and Ideas

  • Symmetry as a Data Augmentation Tool: The paper demonstrates that exploiting object symmetry can effectively "complete" sparse point clouds, providing valuable information that is otherwise missing. This is a powerful way to enhance data without requiring additional sensors or complex fusion techniques.
  • Predicting Symmetry Points: Instead of relying solely on ground truth annotations to generate symmetry points, EPGNet learns to predict their locations, making the approach more generalizable to unseen data.
  • Enhanced Point Cloud for Better Detection: The combination of original points, predicted symmetry points, and voxel centers creates a richer input for the detection backbone, leading to improved performance.
  • Addressing Occlusion and Sparsity: The method shows particular promise in handling occluded and distant objects, where point cloud data is inherently sparse and incomplete.

Key Results

  • Competitive Performance: EPGNet achieves competitive results on the KITTI benchmark for both 3D and BEV object detection, outperforming several state-of-the-art methods, especially in the moderate difficulty level.
  • Improved Detection of Difficult Objects: Qualitative results show EPGNet's ability to detect heavily obscured and far-away cars, and even detect unlabeled cars that the baseline misses.
  • Reduced False Detections: By leveraging symmetry, the network can obtain more structural information, leading to fewer false positive detections compared to the baseline.
  • Better Regression: The generated symmetry points contribute to more accurate 3D box regression, resulting in smaller errors between predicted and ground truth boxes.
  • Efficiency: The method achieves a frame rate of 14 FPS, which is faster than many other two-stage methods.

Limitations and Future Work

  • Segmentation of Sparse Objects: The authors acknowledge that the symmetry point generation is more effective for objects with more points. Objects with fewer points are harder to segment, and their symmetry points might be filtered out if the confidence threshold is too high.
  • Pedestrian Detection: The current method does not support pedestrian detection due to their variable postures, which makes generating accurate symmetry point labels difficult.
  • Future Directions:
    • Developing more effective network structures for segmenting difficult points.
  • Exploring multi-sensor fusion techniques.
  • Investigating domain adaptation methods (e.g., Associate-3Ddet) to further improve generalization.
  • Extending the method to other symmetrical object categories.

Summary updated

BibTeX
@misc{chen2020epgnet,
  author = {Qingsheng Chen and Cien Fan and Weizheng Jin and Lian Zou and Fangyu Li and Xiaopeng Li and Hao Jiang and Minyuan Wu and Yifeng Liu},
  title = {{EPGNet: Enhanced Point Cloud Generation for 3D Object Detection}},
  year = {2020},
  doi = {10.3390/s20236927},
}
Figure