URCDC-Depth: Uncertainty Rectified Cross-Distillation with CutFlip for Monocular Depth Estimation

TL;DR

URCDC-Depth introduces an uncertainty-aware cross-distillation framework combining Transformer and CNN for monocular depth estimation, enhanced by a novel data augmentation technique, CutFlip, achieving state-of-the-art results without extra inference cost.

Contribution

The paper proposes a novel uncertainty rectified cross-distillation method with feature transfer and a simple data augmentation, significantly improving monocular depth estimation accuracy.

Findings

Outperforms previous state-of-the-art on KITTI, NYU-Depth-v2, and SUN RGB-D datasets.

Effectively models pixel-wise depth uncertainty to improve pseudo label quality.

Uses CutFlip augmentation to enhance model robustness and depth inference.

Abstract

This work aims to estimate a high-quality depth map from a single RGB image. Due to the lack of depth clues, making full use of the long-range correlation and the local information is critical for accurate depth estimation. Towards this end, we introduce an uncertainty rectified cross-distillation between Transformer and convolutional neural network (CNN) to learn a unified depth estimator. Specifically, we use the depth estimates from the Transformer branch and the CNN branch as pseudo labels to teach each other. Meanwhile, we model the pixel-wise depth uncertainty to rectify the loss weights of noisy pseudo labels. To avoid the large capacity gap induced by the strong Transformer branch deteriorating the cross-distillation, we transfer the feature maps from Transformer to CNN and design coupling units to assist the weak CNN branch to leverage the transferred features. Furthermore, we propose a surprisingly simple yet highly effective data augmentation technique CutFlip, which enforces the model to exploit more valuable clues apart from the vertical image position for depth inference. Extensive experiments demonstrate that our model, termed~\textbf{URCDC-Depth}, exceeds previous state-of-the-art methods on the KITTI, NYU-Depth-v2 and SUN RGB-D datasets, even with no additional computational burden at inference time. The source code is publicly available at \url{https://github.com/ShuweiShao/URCDC-Depth}.