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 by Fabio Tosi, Luca Bartolomei, Matteo Poggi, and Stefano Mattoccia (arXiv: 2607.08771). While monocular depth foundation models achieve strong zero-shot generalization, their computational demands make them impractical for embedded and mobile platforms; existing lightweight alternatives are mostly single-domain, self-supervised models that fail silently under domain shift. ZipDepth bridges this gap by combining an efficient reparameterizable encoder-decoder with knowledge distillation from a foundation model trained on a large multi-domain dataset. With only 6.1M parameters, ZipDepth runs in real time across hardware ranging from server GPUs to power-constrained devices, achieving the best trade-off between zero-shot accuracy and deployment efficiency among lightweight models on five benchmarks, approaching the accuracy of foundation models roughly 50x its size. This post is a forum repost of the paper abstract, originally auto-collected on 2026-07-12.

Paper Overview

Field: Computer Vision (CV) Authors: Fabio Tosi, Luca Bartolomei, Matteo Poggi, Stefano Mattoccia Published: 2026-07-09 arXiv: 2607.08771

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-model depth estimators generalize well zero-shot but are too heavy for embedded/mobile deployment; existing lightweight models fail under domain shift.
Approach: Efficient reparameterizable encoder-decoder + knowledge distillation from a foundation model on a large multi-domain training set.
Size: Only 6.1M parameters — roughly 50x smaller than the source foundation model.
Performance: Real-time inference from server GPUs down to power-constrained devices; best zero-shot accuracy/efficiency trade-off among lightweight models across five benchmarks.

---

*Auto-collected on 2026-07-12.*