SLURP: Side Learning Uncertainty for Regression Problems

Xuanlong Yu, Gianni Franchi, Emanuel Aldea

Introduction

In safety-critical domains such as autonomous driving, a model's ability to quantify its own uncertainty is as vital as the accuracy of its predictions. While uncertainty estimation is well-understood for classification tasks through discrete probabilities, it remains a significant challenge for continuous regression problems like monocular depth estimation and optical flow. Existing approaches, such as Deep Ensembles or Bayesian methods (e.g., MC-Dropout), often suffer from high computational costs or slow inference times, making them unsuitable for real-time applications.

SLURP (Side Learning Uncertainty for Regression Problems) introduces a lightweight, "plug-and-play" auxiliary network designed to estimate pixel-wise uncertainty for complex regression tasks without requiring modifications to the primary model's architecture.

Methodology: The Side Learner

The core of the SLURP framework is the Side Learner ($g$). Unlike traditional methods that modify the main task model ($f$) to output uncertainty, SLURP keeps the main model frozen. The side learner is trained independently to predict the errors of the main model.

Overview of the SLURP architecture, featuring the main task stream and the side-learning uncertainty estimation stream.

Dual-Input Strategy

The side learner processes two distinct inputs: the original RGB image and the main model's prediction map. This dual-input approach allows the network to perform "semantic recovery." By comparing the raw image features with the predicted output, the side learner can identify discrepancies—such as missing edges or small objects—that signal high uncertainty.

Architecture and Loss Function

The side learner utilizes a multi-scale architecture with DenseNet backbones and a Context Block derived from PWC-net. This block employs dilated convolutions to capture error patterns across varying receptive fields.

A key technical innovation is the training objective. Rather than using standard Mean Squared Error (MSE), which is prone to outlier sensitivity, the authors normalize the prediction error using a tanh function and optimize using Binary Cross Entropy (BCE). Treating the normalized error as a "probability of inaccuracy" leads to more stable convergence and better captures the error distribution across the image.

Key Insights

  • Decoupling: Because the side learner is a separate module, it can be added to any pre-trained regression model. This "plug-and-play" nature means researchers do not need access to the original training code or the ability to modify the primary architecture.
  • Semantic Recovery: The side learner effectively flags areas where the main model fails to capture fine-grained semantic details present in the original input, such as thin structures or complex textures.

Experimental Results

The authors evaluated SLURP across synthetic 1D data, monocular depth estimation (KITTI, Cityscapes), and optical flow (FlyingChairs, Sintel, KITTI 2015).

Qualitative results of SLURP on an optical flow task, showing the input, ground truth, prediction, actual error, and predicted uncertainty.

Performance and Efficiency

SLURP demonstrates a superior balance between accuracy and efficiency:

  • Efficiency: In monocular depth tasks, SLURP achieves a runtime of 88ms, significantly faster than the 286ms required by Deep Ensembles, while utilizing only one-third of the parameters.
  • Accuracy: The method consistently matches or exceeds the performance of Deep Ensembles and MC-Dropout on standard metrics like Area Under Sparsification Error (AUSE) and AUROC.
  • Robustness: Under adverse conditions, such as the "Rainy Cityscapes" benchmark, SLURP maintains a high AUROC (~0.80), reliably identifying unreliable pixels in degraded input data.

Conclusion

SLURP provides a generic and efficient solution for uncertainty quantification in regression. By observing a frozen main model through a side-learning lens, it achieves state-of-the-art results with minimal computational overhead. Future work may extend this concept to other modalities, such as 3D point clouds or audio signals, and further optimize the architecture for low-power embedded systems.

Summary updated

Abstract

It has become critical for deep learning algorithms to quantify their output uncertainties to satisfy reliability constraints and provide accurate results. Uncertainty estimation for regression has received less attention than classification due to the more straightforward standardized output of the latter class of tasks and their high importance. However, regression problems are encountered in a wide range of applications in computer vision. We propose SLURP, a generic approach for regression uncertainty estimation via a side learner that exploits the output and the intermediate representations generated by the main task model. We test SLURP on two critical regression tasks in computer vision: monocular depth and optical flow estimation. In addition, we conduct exhaustive benchmarks comprising transfer to different datasets and the addition of aleatoric noise. The results show that our proposal is generic and readily applicable to various regression problems and has a low computational cost with respect to existing solutions.
BibTeX
@inproceedings{yu2021slurp,
  author = {Xuanlong Yu and Gianni Franchi and Emanuel Aldea},
  title = {{SLURP: Side Learning Uncertainty for Regression Problems}},
  booktitle = {BMVC},
  year = {2021},
  url = {https://openreview.net/forum?id=wOb4fGtDh4},
}
Figure