ZipDepth: Lightweight Zero-Shot Monocular Depth Estimation for Embedded Devices

Forum topic · 小凯 · 2026-07-12

Summary

ZipDepth is a compact monocular depth estimation network presented in arXiv paper 2607.08771 by Fabio Tosi, Luca Bartolomei, Matteo Poggi, and Stefano Mattoccia. While foundation models have enabled robust zero-shot generalization in monocular depth estimation, their computational demands exceed the capabilities of embedded and mobile platforms. Existing lightweight alternatives rely almost exclusively on single-domain, self-supervised training and fail silently under domain shift. ZipDepth bridges this gap by combining an efficient reparameterizable encoder-decoder with large-scale knowledge distillation from a foundation model trained on a multi-domain dataset. With only 6.1M parameters, ZipDepth runs in real time from server GPUs to power-constrained devices and achieves the best trade-off between zero-shot accuracy and deployment efficiency among lightweight models across five benchmarks, approaching the accuracy of foundation models roughly 50 times its size.

Overview

Paper: ZipDepth: Bringing Lightweight Zero-Shot Monocular Depth Anywhere arXiv: 2607.08771 Authors: Fabio Tosi, Luca Bartolomei, Matteo Poggi, Stefano Mattoccia Posted on zhichai.net: 2026-07-12

Original Abstract

Monocular depth estimation has seen remarkable progress through foundation models achieving robust zero-shot generalization, yet their computational demands place them far beyond the reach of embedded and mobile platforms. Lightweight alternatives exist, but have been developed almost exclusively within single-domain, self-supervised paradigms, failing silently under domain shift. We present ZipDepth, a compact monocular depth network that bridges this gap by combining an efficient reparameterizable encoder-decoder with large-scale knowledge distillation from a foundation model over a large multi-domain training set. Comprising just 6.1M parameters, ZipDepth runs at real-time rates from server GPUs to power-constrained devices, achieving the best trade-off between zero-shot accuracy and deployment efficiency.

Key Points

Problem: Foundation models for monocular depth estimation deliver strong zero-shot generalization but are too computationally heavy for embedded and mobile deployment.
Limitation of prior lightweight models: Almost all are built on single-domain, self-supervised paradigms and fail silently when encountering domain shift.
Approach: ZipDepth pairs an efficient reparameterizable encoder-decoder architecture with knowledge distillation from a foundation model over a large multi-domain training set.
Efficiency: Only 6.1M parameters, running at real-time rates from server GPUs down to power-constrained devices.
Results: Best trade-off between zero-shot accuracy and deployment efficiency among lightweight models across five benchmarks — a step toward the accuracy of foundation models roughly 50x its parameter count.