Enhancing Depth Perception Resilience in Dynamic and Adversarial Scenarios

The recent advancements in monocular depth estimation (MDE) and multi-view stereo (MVS) have shown significant strides towards enhancing the robustness and accuracy of depth perception in various environments. Researchers are increasingly focusing on addressing the vulnerabilities and limitations of existing models, particularly in dynamic and adversarial scenarios. Innovations such as adversarial attacks on depth perception, multi-sample refinement techniques, and the activation of self-attention mechanisms in pose regression models are pushing the boundaries of what is possible in MDE. Additionally, the integration of physical constraints and novel network architectures, such as those leveraging dual-pixel images and cross-zone feature propagation, are contributing to more efficient and accurate depth completion methods. These developments collectively indicate a trend towards more resilient and versatile depth estimation systems, capable of handling a wider range of real-world conditions and applications, from autonomous driving to robotics.

Sources

Optical Lens Attack on Monocular Depth Estimation for Autonomous Driving

MultiDepth: Multi-Sample Priors for Refining Monocular Metric Depth Estimations in Indoor Scenes

Activating Self-Attention for Multi-Scene Absolute Pose Regression

A Global Depth-Range-Free Multi-View Stereo Transformer Network with Pose Embedding

Improving Domain Generalization in Self-supervised Monocular Depth Estimation via Stabilized Adversarial Training

PMPNet: Pixel Movement Prediction Network for Monocular Depth Estimation in Dynamic Scenes

CFPNet: Improving Lightweight ToF Depth Completion via Cross-zone Feature Propagation

Revisiting Disparity from Dual-Pixel Images: Physics-Informed Lightweight Depth Estimation

D$^3$epth: Self-Supervised Depth Estimation with Dynamic Mask in Dynamic Scenes

Built with on top of