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

Forum topic · 小凯 · 2026-07-11

Summary

ZipDepth is a compact monocular depth estimation network presented by Fabio Tosi, Luca Bartolomei, and Matteo Poggi (arXiv:2507.08183). While foundation models have enabled strong zero-shot generalization in monocular depth estimation, their computational demands exceed the capabilities of embedded and mobile platforms; existing lightweight models are typically trained in single-domain self-supervised setups 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 over a large multi-domain dataset. With only 6.1M parameters, ZipDepth achieves real-time inference from server GPUs down to power-constrained devices, delivering the best trade-off between zero-shot accuracy and deployment efficiency among lightweight models across five benchmarks, approaching the accuracy of foundation models 50x its size.

Overview

Field: Computer Vision Authors: Fabio Tosi, Luca Bartolomei, Matteo Poggi arXiv: 2507.08183

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

Combines a reparameterizable encoder-decoder with knowledge distillation from a foundation model over multi-domain training data.
Only 6.1M parameters; real-time performance across devices from server GPUs to power-constrained embedded hardware.
Best zero-shot accuracy vs. deployment efficiency trade-off among lightweight models across five benchmarks, moving closer to foundation models 50x larger in parameter count.