FinderNet: A Data Augmentation Free Canonicalization Aided Loop Detection and Closure Technique for Point Clouds in 6-DOF Separation.

Sudarshan S. Harithas, Gurkirat Singh, Aneesh Chavan, Sarthak Sharma, Suraj Patni, Chetan Arora, Madhava KrishnaWACV 2024

TL;DR

This paper introduces FinderNet, a novel approach for loop detection and closure in 3D point clouds that significantly reduces the need for data augmentation. By transforming point clouds into "digital elevation maps" (DEMs) and employing a unique yaw transformer, FinderNet achieves robust 6-DOF pose estimation and outperforms state-of-the-art methods on challenging datasets.

What Problem Does This Paper Address?

Mobile robots and autonomous vehicles rely heavily on Simultaneous Localization and Mapping (SLAM) to navigate their environments. A critical component of SLAM is loop detection and closure (LDC), which allows the system to recognize previously visited locations and correct accumulated drift in its pose estimation. Traditional LDC methods often struggle with significant viewpoint variations (6-DOF, meaning changes in position and orientation) and require extensive data augmentation during training to achieve robustness. This data augmentation is computationally expensive and doesn't always generalize well to unseen scenarios. Furthermore, raw 3D point clouds lack inherent structure, making it difficult for algorithms to extract reliable features for accurate LDC.

Methods

FinderNet tackles these challenges with a multi-stage approach:

  • Canonicalization: The core idea is to create a more structured representation of the point cloud. First, the system identifies a dominant ground plane and uses it to canonicalize the point cloud by correcting for "roll" and "pitch" (rotations around the x and y axes). This is achieved through a combination of coarse estimation and Iterative Closest Point (ICP) refinement.
  • DEM Generation: The roll-and-pitch canonicalized point clouds are then transformed into a "Digital Elevation Map" (DEM). This is done by discretizing the point cloud into a grid and storing the maximum height within each grid cell. This DEM representation effectively exposes strong spatial priors and is analogous to a top-down view of the terrain and objects.
  • Autoencoder: A standard autoencoder-decoder architecture is used to learn a latent representation of these DEMs. The encoder compresses the DEM into a lower-dimensional latent space, and the decoder reconstructs the DEM from this latent representation.
  • Differentiable Yaw Transformer (DYT): To handle the remaining "yaw" (rotation around the z-axis) variations, FinderNet introduces a novel Differentiable Yaw Transformer. This module takes the latent embeddings of two point clouds (an anchor and a potential match) and learns to align them rotationally without requiring explicit yaw supervision. This is a key innovation that enables data-augmentation-free training for yaw invariance.
  • Loop Detection and Closure: After yaw alignment, a difference layer computes the absolute differences between the feature volumes of the aligned DEMs. This difference is then fed into a CNN to produce a scalar value indicating the similarity between the two point clouds. A triplet loss is used to train the loop detection module. For loop closure, the system uses the decoded DEMs to estimate the full 6-DOF relative pose.

Datasets

The paper utilizes several datasets for training and evaluation:

  • KITTI [9]: A widely used dataset for autonomous driving research, including sequences for training and testing LDC.
  • GPR [32]: A dataset for long-term autonomy, used for evaluating LDC performance.
  • Lidar UrbanFly Dataset (LUF): A custom dataset created by the authors using Unreal Engine. This dataset features a 64-channel LiDAR mounted on a quadrotor, capturing 6-DOF motion in a simulated urban environment. It's designed to specifically test robustness to significant viewpoint changes.

Key Insights and Ideas

  • DEMs as Structured Representations: The transformation of raw point clouds into DEMs is a crucial insight. It effectively converts unstructured 3D data into a structured 2D representation that is more amenable to standard deep learning architectures like CNNs, while preserving essential spatial information.
  • Data Augmentation is Overrated (for Yaw): The DYT module's ability to achieve yaw invariance without explicit data augmentation is a significant breakthrough. This drastically simplifies the training process and improves generalization.
  • Canonicalization for Robustness: The initial roll and pitch canonicalization, by aligning point clouds to a common ground plane, provides a strong prior that makes subsequent processing more robust to large viewpoint changes.
  • End-to-End 6-DOF LDC: The paper demonstrates a complete pipeline for estimating the full 6-DOF relative pose, which is a more challenging and practical problem than just estimating 2D (SE(2)) transformations.

Key Results

FinderNet demonstrates significant improvements over state-of-the-art methods:

  • Loop Detection: On the KITTI08 sequence, FinderNet achieved an 8.4% improvement in Average Precision (AP) compared to LCDNet. On the GPR10 sequence, it showed an 11% improvement. On the custom LUF dataset, FinderNet outperformed LCDNet by 10% AP.
  • Loop Closure: When integrated with LIO-SAM (a SLAM system), FinderNet reduced the RMSE of translation error by 16% on the challenging KITTI08 sequence compared to LIO-SAM's built-in LDC.
  • Data Efficiency: The method does not require data augmentation for training, which is a major advantage.
  • Robustness: The paper highlights FinderNet's robustness to large viewpoint changes, as demonstrated on the LUF dataset and challenging sequences like KITTI08.

Limitations and Future Work

  • Ground Plane Assumption: The method relies on the assumption of a dominant ground plane, which might not hold in all environments (e.g., aerial scenes without a clear ground).
  • Dynamic Scenes: The current work focuses on static scenes. The authors suggest that extending the method to handle dynamic scenes is a promising future direction.
  • Real-time Performance: While the paper mentions integration with LIO-SAM for real-time efficacy, further exploration of optimizing the pipeline for high-speed applications could be beneficial.

Summary updated

BibTeX
@inproceedings{harithas2024findernet,
  author = {Sudarshan S. Harithas and Gurkirat Singh and Aneesh Chavan and Sarthak Sharma and Suraj Patni and Chetan Arora and Madhava Krishna},
  title = {{FinderNet: A Data Augmentation Free Canonicalization Aided Loop Detection and Closure Technique for Point Clouds in 6-DOF Separation.}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
}
Figure