Enhancing Unsupervised Monocular Depth Estimation for Zero-Shot Domain Adaptation Across Smartphone Cameras
Year: 2025
Type: conference-paper
Abstract: Accurate monocular depth estimation is essential for applications such as autonomous navigation and augmentedreality. However, most models require large-scale labeled datasets, limiting their adaptability across diverseenvironments. Inspired by the SPIdepth model (Lavreniuk 2024), which achieved state-of-the-art depth estimationaccuracy on the Cityscape and Kitti benchmarks (Cordts et al. 2016; Geiger et al. 2013) in 2024, this study proposesan enhanced unsupervised monocular depth estimation framework that improves zero-shot generalization acrossdifferent smartphone cameras without relying on ground truth depth maps .Our analysis of SPIdepth reveals a lack of domain adaptation, as shown in the cross -domain testing results inTable 1. Models trained on KITTI (K) or Cityscapes (C) perform well within their respective domains but degradesignificantly when tested on unseen datasets. For instance, a model trained on K and fine-tuned on C (K→C) performswell on KITTI (AbsRel = 0.067, RMSE = 3.379) but suffers on Cityscapes (AbsRel = 0.123, RMSE = 6.837).Similarly, a model trained on C and fine-tuned on K (C→K) shows an AbsRel of 0.083 on Cityscapes but degradesto 0.13 on KITTI. Notably, when models are trained and tested on the same dataset, their performance improves,highlighting the importance of domain-specific training. Training on Cityscapes and testing on Cityscapes results ina lower AbsRel (0.083) and RMSE (5.207) compared to training on KITTI, while training on KITTI and testing onKITTI achieves the best performance (AbsRel = 0.067, RMSE = 3.379). These results emphasize the limitedgeneralization of the current method beyond their training datasets and underscore the need for improved domainadaptation techniques.
Institution: University of Calgary
Language: English
Topic: Advanced Vision and Imaging
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213622985