Abstract:Uncertainty-aware semantic segmentation of the point clouds includes predictive uncertainty estimation and uncertainty-guided model optimization. One key challenge in the task is the efficiency of pointwise predictive distribution establishment. The widely used Monte Carlo (MC) dropout establishes the distribution by computing the standard deviation of samples using multiple stochastic forward propagations, which is time-consuming for tasks based on point clouds containing massive points. Hence, a framework embedded with neighborhood spatial aggregation (NSA)-MC dropout, a variant of MC dropout, is proposed to establish distributions in just one forward pass. Specifically, our method uses the one-time stochastic inference of a point with neighbors to approximate the point’s repeated stochastic inferences, outputting pointwise distribution via the prediction variance of neighbors. Based on this, uncertainties acquire from the predictive distribution. The aleatoric uncertainty is integrated into the loss function to suppress noise, preventing the model from overfitting. Besides, the predictive uncertainty quantifies the prediction confidence. Experiments show that our plug-and-play NSA-MC dropout significantly improves the segmentation performance of backbones, ranging from multilayer perceptron (MLP)-, convolution-, and attention-based to transformer-based networks, without introducing additional parameters. Besides, it is several times faster than MC dropout to quantify the results’ credibility, and the inference time does not establish a coupling relation with the sampling times.

Revisiting Softmax for Uncertainty Approximation in Text Classification

Understanding Softmax Confidence and Uncertainty

Towards More Accurate Uncertainty Estimation in Text Classification

Inhibited Softmax for Uncertainty Estimation in Neural Networks

Illustrative Discussion of MC-Dropout in General Dataset: Uncertainty Estimation in Bitcoin

Density-Softmax: Efficient Test-time Model for Uncertainty Estimation and Robustness under Distribution Shifts

Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning

Open Set Recognition Through Monte Carlo Dropout-Based Uncertainty.

Applying Monte Carlo Dropout to Quantify the Uncertainty of Skip Connection-Based Convolutional Neural Networks Optimized by Big Data

Efficient Uncertainty Quantification for Multilabel Text Classification

Single-shot Bayesian approximation for neural networks

A Novel Regression Loss for Non-Parametric Uncertainty Optimization

Efficient Uncertainty Estimation in Spiking Neural Networks via MC-dropout

Uncertainty estimation in Deep Learning for Panoptic segmentation

Energy-Based Open-World Uncertainty Modeling for Confidence Calibration

Uncertainty propagation for dropout-based Bayesian neural networks

Comparing and Improving Active Learning Uncertainty Measures for Transformer Models by Discarding Outliers

On Last-Layer Algorithms for Classification: Decoupling Representation from Uncertainty Estimation

What Does Softmax Probability Tell Us about Classifiers Ranking Across Diverse Test Conditions?

Neighborhood Spatial Aggregation MC Dropout for Efficient Uncertainty-aware Semantic Segmentation in Point Clouds

The Peril of Popular Deep Learning Uncertainty Estimation Methods