MonoPP: Metric-Scaled Self-Supervised Monocular Depth Estimation by Planar-Parallax Geometry in Automotive Applications

Gasser Elazab, Torben Gräber, Michael Unterreiner, Olaf HellwichWACV 2025

TL;DR

This paper introduces MonoPP, a novel self-supervised monocular depth estimation method for automotive applications that overcomes the scale-ambiguity of existing approaches. By leveraging planar-parallax geometry and the readily available camera mounting position, MonoPP achieves metric-scaled depth predictions without requiring ground-truth data, demonstrating state-of-the-art performance on the KITTI benchmark and showing promise for challenging datasets like Cityscapes.

What Problem Does This Paper Address?

Monocular Depth Estimation (MDE) is crucial for autonomous vehicles to understand their surroundings, enabling safe navigation and obstacle detection. While deep learning has significantly advanced MDE, many self-supervised methods produce scale-invariant depth maps, meaning they lack absolute metric information (e.g., depth in meters). This scale ambiguity limits their practical use in real-world scenarios where precise depth measurements are essential. Supervised methods typically rely on expensive sensors like LiDAR or RGB-D cameras for ground truth, which are not always available. This paper addresses the challenge of achieving metric-scaled depth estimation using only monocular video data and the camera's mounting position, a common configuration in modern vehicles.

Methods

MonoPP employs a two-pipeline architecture: a single-frame network for direct depth prediction and a multi-frame "teacher" network that leverages planar-parallax geometry.

  • Planar-Parallax Geometry: This core technique reconstructs scene structure by analyzing the apparent motion (residual flow) of static objects relative to a planar surface (like the road). It decomposes scene changes into planar homography and epipolar flow.
  • Teacher-Student Distillation: The multi-frame teacher network, using planar-parallax geometry and the camera's mounting height, reconstructs the static scene structure and predicts metric-scaled depth. This teacher then "distills" knowledge, including scale information, a static mask, and metric-scaled depth, to a simpler single-frame student network.
  • Pose Network: A ResNet-18-based pose network estimates the relative rotation and translation between sequential frames, which is crucial for computing homography and enabling the planar-parallax calculations.
  • Novelty in Scaling: Unlike many self-supervised methods that rely on per-frame median ground truth for scaling, MonoPP directly infers metric scale using the camera's height above the ground and the planar-parallax geometry.
  • Masking Dynamic Objects: The teacher network also generates a static mask to filter out dynamic objects, preventing them from corrupting the training signals for the depth estimation.
  • Loss Functions: The training involves several loss functions, including photometric reprojection loss, homography loss (for road alignment and metric scaling), planar-parallax depth loss, and a consistency loss to ensure agreement between the teacher and student predictions.

Datasets

  • KITTI: Used for training and evaluation, specifically the Eigen split, a standard benchmark for single-frame depth estimation.
  • Cityscapes: Used for evaluation, demonstrating the model's effectiveness on a more challenging dataset where self-supervised metric-scaled depth prediction is less common.

Key Insights and Ideas

  • Camera Height as a Powerful Cue: The paper demonstrates that the camera's mounting height above the ground, a readily available piece of information in automotive settings, is a sufficient cue for inferring metric scale in self-supervised depth estimation.
  • Planar-Parallax for Structure and Scale: Planar-parallax geometry proves to be an effective tool not only for reconstructing static scene structure but also for guiding the estimation of metric scale, especially when dealing with planar surfaces like roads.
  • Teacher-Student Distillation for Efficiency: Using a computationally expensive teacher network to distill knowledge into a lighter student network allows for achieving high performance while enabling potentially faster inference with the student model.
  • Effective Dynamic Object Handling: The explicit generation of a static mask by the teacher network is crucial for mitigating the negative impact of moving objects on self-supervised training.

Key Results

  • State-of-the-Art on KITTI: MonoPP achieved state-of-the-art results for metric-scaled depth prediction on the KITTI benchmark.
  • First Self-Supervised Metric-Scaled Depth on Cityscapes: The paper reports being among the first to achieve self-supervised metric-scaled depth prediction on the challenging Cityscapes dataset, showcasing its versatility.
  • Competitive Performance: MonoPP's metric-depth estimation results are competitive with methods that utilize pose information, while MonoPP only requires camera height.
  • Inference Speed: The inference model exhibits a latency of approximately 0.019 seconds per image on an Nvidia T4 GPU, indicating good real-time potential.

Limitations and Future Work

  • Real-time Performance in Vehicle: The authors plan to evaluate the real-time performance of their model in an actual vehicle, which will involve using continuous camera mounting position data and accounting for road slopes.
  • Assumption of Planar Surfaces: While planar-parallax geometry is powerful, its effectiveness might be reduced in scenes with limited or no dominant planar surfaces.
  • Dynamic Object Handling: While masked, the presence of highly dynamic or complex dynamic scenes could still pose challenges.

Summary updated

BibTeX
@inproceedings{elazab2025monopp,
  author = {Gasser Elazab and Torben Gräber and Michael Unterreiner and Olaf Hellwich},
  title = {{MonoPP: Metric-Scaled Self-Supervised Monocular Depth Estimation by Planar-Parallax Geometry in Automotive Applications}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure