Zero-Shot Synth-to-Real Depth Estimation: From Synthetic Street Scenes to Real-World Data

This paper introduces a novel method for estimating depth maps from single images using a convolutional neural network (CNN) architecture. Our approach leverages synthetically generated data that simulates front views from autonomous vehicles. The model employs a ConvNeXt encoder and a U-Net-based decoder, achieving effective depth estimation performance. A scale- and shift-invariant loss function is utilized during training to enhance generalization capabilities. The proposed model demonstrates strong results on real-world datasets without requiring fine-tuning, highlighting the effective transfer of information from synthetic to real-world data. Additionally, we show that high-quality data significantly improves performance compared to larger, low-quality datasets.