[Paper Review] GANav: Group-wise Attention Network for Classifying Navigable Regions in Unstructured Outdoor Environments.

Tianrui Guan, Divya Kothandaraman+2 more|arXiv (Cornell University)|Mar 7, 2021

Advanced Neural Network Applications37 references9 citations

TL;DR

GANav proposes a group-wise attention transformer network for classifying navigable regions in unstructured outdoor environments using RGB images. By leveraging group-wise attention and a dynamic weighted cross-entropy loss, it improves navigability segmentation accuracy, achieving 6.74–39.1% higher mIoU on RUGD and 3.82–10.64% on RELLIS-3D compared to state-of-the-art methods.

ABSTRACT

We present a new learning-based method for identifying safe and navigable regions in off-road terrains and unstructured environments from RGB images. Our approach consists of classifying groups of terrain classes based on their navigability levels using coarse-grained semantic segmentation. We propose a bottleneck transformer-based deep neural network architecture that uses a novel group-wise attention mechanism to distinguish between navigability levels of different terrains.Our group-wise attention heads enable the network to explicitly focus on the different groups and improve the accuracy. In addition, we propose a dynamic weighted cross entropy loss function to handle the long-tailed nature of the dataset. We show through extensive evaluations on the RUGD and RELLIS-3D datasets that our learning algorithm improves the accuracy of visual perception in off-road terrains for navigation. We compare our approach with prior work on these datasets and achieve an improvement over the state-of-the-art mIoU by 6.74-39.1% on RUGD and 3.82-10.64% on RELLIS-3D.

Motivation & Objective

• To improve visual perception of navigable regions in unstructured outdoor terrains using RGB images.
• To address the long-tailed distribution of terrain classes in off-road datasets.
• To enhance segmentation accuracy by explicitly modeling navigability levels across terrain groups.
• To develop a deep learning architecture that captures hierarchical relationships between terrain classes.

Proposed method

• Proposes a bottleneck transformer-based neural network with a novel group-wise attention mechanism to focus on distinct terrain groups.
• Designs a dynamic weighted cross-entropy loss function to mitigate the impact of long-tailed class distributions in training data.
• Uses coarse-grained semantic segmentation to classify terrain regions based on navigability levels.
• Groups terrain classes into navigability levels to guide attention and improve generalization.
• Integrates the group-wise attention heads to explicitly attend to different terrain groupings during feature representation learning.
• Trains and evaluates the model on the RUGD and RELLIS-3D datasets to benchmark performance on real-world off-road scenarios.

Experimental results

Research questions

• RQ1 Can a group-wise attention mechanism improve the accuracy of navigability classification in unstructured outdoor environments?
• RQ2 How does the dynamic weighted cross-entropy loss affect model performance on long-tailed terrain datasets?
• RQ3 To what extent does the proposed architecture outperform existing state-of-the-art methods on RUGD and RELLIS-3D?
• RQ4 Does modeling terrain classes in navigability groups lead to better generalization and robustness in segmentation?

Key findings

• GANav achieves a 6.74–39.1% improvement in mean Intersection over Union (mIoU) on the RUGD dataset compared to state-of-the-art methods.
• On the RELLIS-3D dataset, the model improves mIoU by 3.82–10.64% over prior approaches.
• The group-wise attention mechanism enhances feature discrimination by explicitly focusing on distinct terrain groups.
• The dynamic weighted cross-entropy loss effectively mitigates performance degradation due to long-tailed class distributions.
• The model demonstrates strong generalization across diverse off-road environments, as validated on two benchmark datasets.
• The proposed architecture achieves state-of-the-art performance in visual perception for off-road navigation using only RGB input.