PromptMono: Cross Prompting Attention for Self-Supervised Monocular Depth Estimation in Challenging Environments

TL;DR

PromptMono tackles the challenge of monocular depth estimation in diverse and adverse environments by introducing visual prompt learning and a gated cross prompting attention mechanism. A siamese prompting-based learning framework, augmented with self-distillation and image-to-image domain translation, enables a single unified model to predict depth across day, night, and rain conditions without domain-specific branches. The GCPA module fuses learnable prompts with image features through CGPB and multi-head cross-attention, and a prompting-based depth decoder integrates this information into multi-scale depth maps. Empirical results on Oxford RobotCar and nuScenes show consistent improvements over strong self-supervised baselines, especially under challenging illumination and weather, highlighting the practical value of cross-domain prompting for depth perception. The approach offers a flexible, hardware-efficient path to robust monocular depth estimation in real-world autonomous systems and can be extended to other vision tasks requiring cross-domain generalization.

Abstract

Considerable efforts have been made to improve monocular depth estimation under ideal conditions. However, in challenging environments, monocular depth estimation still faces difficulties. In this paper, we introduce visual prompt learning for predicting depth across different environments within a unified model, and present a self-supervised learning framework called PromptMono. It employs a set of learnable parameters as visual prompts to capture domain-specific knowledge. To integrate prompting information into image representations, a novel gated cross prompting attention (GCPA) module is proposed, which enhances the depth estimation in diverse conditions. We evaluate the proposed PromptMono on the Oxford Robotcar dataset and the nuScenes dataset. Experimental results demonstrate the superior performance of the proposed method.

Figures (5)

• Figure 1: Depth estimation in challenging environments. The first column and the second column are examples from the Oxford Robotcar dataset robotcar and the nuScenes dataset nuscenes, respectively. Low visibility at night and reflections on wet roads during rain are detrimental to depth estimation as depicted in the second row.
• Figure 2: An overview of the proposed PromptMono framework. It leverages visual prompt learning with a self-distillation scheme to train a unified model for predicting depth in diverse environments.
• Figure 3: An overview of the proposed GCPA module. Taking image feature and visual prompts as inputs, the GCPA module first processes the prompts using the proposed CGPB, and then captures domain-specific information by computing the multi-head cross-attention between the input features and the prompts. Finally, it outputs the prompt-enhanced features using a gated-DConv feed-forward network restormer/cvpr22.
• Figure 4: Qualitative results on the Oxford RobotCar dataset.
• Figure 5: Qualitative results on the nuScenes dataset.