Robust Multi-view Depth Benchmark

Robust Multi-view Depth Benchmark是一个基于多个公共数据集构建的基准，旨在评估多视角深度估计模型在不同领域数据上的鲁棒性。该基准支持视频深度获取和多视角立体设置下的评估，并特别考虑了相机姿态已知时的绝对深度估计。数据集涵盖了多个领域，包括驾驶、室内外场景等，用于零样本跨数据集测试，无需微调。此基准还提供了一个基于现有组件的基准模型，该模型通过新颖的尺度增强程序，能够在不同目标数据上实现鲁棒的多视角深度估计。Robust Multi-view Depth Benchmark is a benchmark constructed upon multiple public datasets, designed to evaluate the robustness of multi-view depth estimation models across diverse domain data. This benchmark supports evaluations under both video depth acquisition and multi-view stereo settings, with special focus on absolute depth estimation when camera poses are known. The dataset covers various domains including driving scenarios, indoor and outdoor environments, etc., enabling zero-shot cross-dataset testing without fine-tuning. Additionally, this benchmark provides a baseline model built from existing components, which achieves robust multi-view depth estimation across different target datasets via a novel scale-augmentation pipeline.

搜集汇总

背景与挑战

背景概述

Robust Multi-view Depth Benchmark是一个多领域深度估计基准，用于评估模型在不同数据上的鲁棒性，支持视频和多视角立体设置，并包含零样本跨数据集测试。它提供了一个基准模型，通过尺度增强实现鲁棒的深度估计，适用于驾驶、室内外等场景，无需微调。