[Paper Review] Uncertainty-aware Mean Teacher for Source-free Unsupervised Domain Adaptive 3D Object Detection

Deepti Hegde, Vishwanath A. Sindagi+4 more|arXiv (Cornell University)|Sep 29, 2021

Advanced Neural Network ApplicationsComputer Science32 references19 citations

TL;DR

This paper proposes an uncertainty-aware mean teacher framework for source-free unsupervised domain adaptation in 3D object detection, leveraging Monte Carlo dropout to estimate model uncertainty and down-weight losses from unreliable pseudo-labels during training. The method achieves state-of-the-art performance on cross-dataset and cross-weather domain shifts, outperforming prior SOTA methods like ST3D and semi-supervised baselines on KITTI, Waymo, and nuScenes datasets.

ABSTRACT

Pseudo-label based self training approaches are a popular method for source-free unsupervised domain adaptation. However, their efficacy depends on the quality of the labels generated by the source trained model. These labels may be incorrect with high confidence, rendering thresholding methods ineffective. In order to avoid reinforcing errors caused by label noise, we propose an uncertainty-aware mean teacher framework which implicitly filters incorrect pseudo-labels during training. Leveraging model uncertainty allows the mean teacher network to perform implicit filtering by down-weighing losses corresponding uncertain pseudo-labels. Effectively, we perform automatic soft-sampling of pseudo-labeled data while aligning predictions from the student and teacher networks. We demonstrate our method on several domain adaptation scenarios, from cross-dataset to cross-weather conditions, and achieve state-of-the-art performance in these cases, on the KITTI lidar target dataset.

Motivation & Objective

• To address domain shift in 3D object detection caused by differences in lidar sensor types, environments, and weather conditions between source and target domains.
• To overcome the limitations of pseudo-label based self-training in source-free unsupervised domain adaptation, where high-confidence incorrect labels can degrade model performance.
• To develop a method that improves detection accuracy without access to source domain data or labels, enabling practical deployment in privacy-sensitive or data-restricted scenarios.
• To implicitly filter noisy pseudo-labels during training by leveraging model uncertainty, avoiding reliance on confidence thresholds alone.

Proposed method

• Uses a mean teacher framework with a student and teacher network, both initialized from a source-trained model (e.g., SECOND-IoU), where the teacher’s weights are slowly updated via exponential moving average.
• Generates pseudo-labels for the target domain iteratively using the student network, refining them over multiple training stages to improve label quality.
• Employs Monte Carlo dropout to estimate model uncertainty by performing T forward passes through the teacher network, computing variance per prediction to quantify uncertainty.
• Applies uncertainty-aware loss weighting: samples with high uncertainty (high variance) are down-weighted during training, reducing the impact of noisy or incorrect pseudo-labels.
• Implements soft-sampling by dynamically adjusting loss contributions based on uncertainty, enabling implicit filtering of low-quality pseudo-labeled examples.
• Trains the student network using a combination of ground-truth labels from the source and pseudo-labeled targets, with uncertainty-aware supervision from the teacher network.

Experimental results

Research questions

• RQ1 Can uncertainty estimation via Monte Carlo dropout improve the robustness of pseudo-label based self-training in source-free domain adaptation for 3D object detection?
• RQ2 How does uncertainty-aware loss weighting compare to threshold-based filtering in mitigating the impact of high-confidence noisy pseudo-labels?
• RQ3 To what extent does iterative pseudo-label refinement enhance detection performance across cross-dataset and adverse weather domain shifts?
• RQ4 Does the proposed method outperform existing source-free and semi-supervised domain adaptation approaches on benchmark 3D detection datasets like KITTI, Waymo, and nuScenes?

Key findings

• The proposed method achieves state-of-the-art mean average precision (mAP) across all evaluated domain shift scenarios, including Waymo → KITTI, Waymo → nuScenes, and cross-weather conditions.
• On the KITTI dataset, the method outperforms ST3D and the semi-supervised statistical normalization baseline, achieving higher mAP in all difficulty categories.
• The uncertainty-aware mean teacher framework shows significant performance gains over non-uncertainty-aware variants, especially in reducing false positives from high-confidence incorrect pseudo-labels.
• Iterative pseudo-label generation improves detection mAP up to a point (J ≈ 3 iterations), after which performance plateaus or declines, indicating optimal refinement stages.
• Model uncertainty analysis confirms that by the end of training, the teacher network assigns higher variance to incorrect pseudo-labels, enabling effective down-weighting during optimization.
• Qualitative results show that the method produces more accurate and localized detections than both the source-only model and ST3D, particularly in reducing false positives with high confidence.