MonoMixer: Marrying Convolution and Vision Transformer for Efficient Self-Supervised Monocular Depth Estimation

Zhiyong Chang, Yan WangIJCAI 2025 · DOI

Abstract

Self-supervised monocular depth estimation (MDE) is a critical task for autonomous systems, yet existing models often struggle to balance computational efficiency with the need for global scene context. This paper introduces MonoMixer, an ultra-lightweight hybrid architecture that integrates the local precision of Convolutional Neural Networks (CNNs) with the global reasoning of Vision Transformers (ViTs). By replacing standard, computationally expensive self-attention with a novel global-guided cross-attention mechanism, MonoMixer achieves state-of-the-art accuracy while utilizing significantly fewer parameters than previous leading models.

Introduction

Monocular Depth Estimation (MDE)—the process of predicting 3D depth from a single 2D image—is a cornerstone of autonomous driving and robotics. Due to the high cost of gathering ground-truth depth data via LiDAR, self-supervised learning has emerged as a preferred paradigm. In this approach, models learn to infer depth by reconstructing one view of a scene from another using geometric constraints.

However, current MDE architectures face a significant trade-off. CNN-based models are computationally efficient and excel at capturing local details but lack the receptive field necessary to understand global scene layouts. Conversely, Transformer-based models (ViTs) capture global relationships effectively but suffer from quadratic computational complexity, resulting in massive parameter counts that hinder deployment on edge devices like drones or mobile robots. MonoMixer addresses this "Goldilocks" problem by combining a global "brain" with a lightweight "body."

Methodology: The MonoMixer Architecture

MonoMixer consists of an encoder-decoder "DepthNet" and a "PoseNet." The core innovation lies in three specialized blocks within the encoder designed to maximize information flow while minimizing overhead.

The MonoMixer architecture features a DepthNet with novel G2T, DA, and SMCA blocks, alongside a PoseNet for motion estimation.

Global-Guided Transformer (G2T) Block

Standard self-attention requires every pixel to attend to every other pixel, leading to prohibitive costs. MonoMixer introduces a Global Semantic Token that summarizes the entire image. Local features then perform cross-attention only with this single token. This strategy reduces computational complexity from quadratic to linear while ensuring that global context is injected into every local feature.

Details-Augmented (DA) Block

CNNs often produce blurred boundaries for thin objects. The DA block utilizes Graph Convolutional Networks (GCNs) to model the topological structure of the scene. By treating pixel clusters as nodes in a graph, the model can reason more effectively about object edges and slender structures, such as poles or signs, which are often lost in traditional downsampling.

Self-Modulate Channel Attention (SMCA) Block

Inspired by the MLP-Mixer, the SMCA block adaptively weights feature channels. This mechanism helps the model emphasize critical geometric patterns and suppress noise, further refining the resulting depth map.

Experimental Results

The model was evaluated on the KITTI, Make3D, and Cityscapes benchmarks.

Quantitative Performance

MonoMixer demonstrates a remarkable efficiency-to-accuracy ratio. As shown in the table below, it achieves an Absolute Relative Error (AbsRel) of 0.081 on the KITTI dataset, outperforming models with over ten times its parameter count.

Quantitative comparison on the KITTI benchmark. MonoMixer (Ours) achieves state-of-the-art performance with only 3.9M parameters.

Efficiency and Real-Time Deployment

With only 3.9 million parameters, MonoMixer is significantly smaller than MonoFormer (23.9M) and SQLdepth (34M). This lightweight design translates to high-speed inference: the model runs in 2.6ms on a Titan XP and 10.5ms on a Jetson Xavier edge module, making it highly suitable for real-world robotics applications.

Qualitative Analysis

Visual comparisons reveal that MonoMixer excels at capturing "slender structures." While traffic light poles and distant road signs often disappear or blur in other models' depth maps, MonoMixer maintains sharp boundaries and accurate depth values for these challenging features.

Discussion and Future Work

MonoMixer proves that architectural intelligence—specifically the strategic combination of local and global processing—can outperform raw scaling. However, challenges remain. Like most self-supervised MDE models, MonoMixer assumes a static scene, which can lead to errors in highly congested urban environments. Additionally, the model currently relies on median scaling for metric depth, and future work will explore learning absolute scale without labels. Finally, while inference is fast, further optimization for ultra-low-power microcontrollers remains an open area for research.

Conclusion

MonoMixer represents a significant step forward in efficient depth estimation. By "marrying" convolutions and transformers through global-guided attention and graph-based detail augmentation, it provides a high-performance solution that is finally viable for deployment on resource-constrained edge devices.

Summary updated

BibTeX
@inproceedings{chang2025monomixer,
  author = {Zhiyong Chang and Yan Wang},
  title = {{MonoMixer: Marrying Convolution and Vision Transformer for Efficient Self-Supervised Monocular Depth Estimation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/85},
}
Figure