Coarse-to-Fine Self-supervised Monocular Depth Estimation of Dynamic Objects with Ground Contact Prior

This page is under reconstruction
Corresponding author

Input Image
Input Image
Input Image
Input Image


Self-supervised monocular depth estimation (DE) is an approach to learning depth without costly depth ground truths. However, it often struggles with moving objects that violate the static scene assumption during training. To address this issue, we introduce a coarse-to-fine training strategy leveraging the ground contacting prior based on the observation that most moving objects in outdoor scenes contact the ground. In the coarse training stage, we exclude the objects in dynamic classes from the reprojection loss calculation to avoid inaccurate depth learning. To provide precise supervision on the depth of the objects, we present a novel Ground-contacting-prior Disparity Smoothness Loss (GDS-Loss) that encourages a DE network to align the depth of the objects with their ground-contacting points. Subsequently, in the fine training stage, we refine the DE network to learn the detailed depth of the objects from the reprojection loss, while ensuring accurate DE on the moving object regions by employing our regularization loss with a cost-volume-based weighting factor. Our overall coarse-to-fine training strategy can easily be integrated with existing DE methods without any modifications, significantly enhancing DE performance on challenging Cityscapes and KITTI datasets, especially in the moving object regions.


Method Overview

We propose a novel coarse-to-fine training strategy to effectively handle moving object problems in self-supervised learning of monocular depth estimation.
In the coarse training stage, we utilize the ground contacting prior as a self-supervision for depth estimation on dynamic objects, presenting Ground-contacting-prior Disparity Smoonthess Loss (GDS-Loss). The ground contacting prior is based on an observation that most objects classified as dynamic in outdoor scenes, such as cars, bicycles, or pedestrians, invariably tend to make contact with the ground, thereby sharing similar depth at their ground contact points.
In the fine training stage, we further refine the depth estimation network to capture detailed depth of dynamic objects. We introduce a regularization loss with a cost-volume-based weighting factor to avoid inaccurate learning from the reproejction loss on the moving object regions.

Graph for quality evaluation

SOTA Comparison

Our coarse-to-fine training strategy significantly improves the depth estimation performance of existing methods including Monodepth2 [ICCV 2019], HR-Depth [AAAI 2021], CADepth [3DV 2021] and MonoViT [3DV 2022]. Each model trained with our training strategy is denoted as Ours-Monodepth2, Ours-HR-Depth, Ours-CADepth, Ours-MonoViT. Our training strategy is easily integrated into those methods and enhances depth estimation performance on both Cityscapes and KITTI datasets. Moreover, Ours-MonoViT achieves state-of-the-art depth estimation performance on both datasets.

Graph for quality evaluation

3D Point Cloud Reconstructions

Graph for quality evaluation


      title={From-Ground-To-Objects: Coarse-to-Fine Self-supervised Monocular Depth Estimation of Dynamic Objects with Ground Contact Prior},
      author={Moon, Jaeho and Bello, Juan Luis Gonzalez and Kwon, Byeongjun and Kim, Munchurl},
      journal={arXiv preprint arXiv:2312.10118},