ArchDepth
Depth estimation from 2D images is a fundamental research topic in photogrammetry and computer vision toward 3D reconstruction and scene understanding with a vast field of applications, including mapping, navigation, and augmented reality. A recent trend focuses on learning-based monocular depth estimation algorithms and despite the release of large-scale datasets for training, methods are limited in scenarios and still struggle to yield reliable results in the 3D space without additional scene cues. Moreover the transferability of deep learning depth estimation for real-world photogrammetric scenarios is still a challenging problem. We therefore investigated the potential of integrating learning-based monocular depth estimation in photogrammetric applications.
Our contribution includes:
(1) a novel rendered large-scale dataset (ArchDepth) of photorealistic outdoor scenes of historic buildings, including high-quality, complete, metric depth maps for every image;
(2) a straightforward training pipeline following an encoder-decoder network for metric monocular depth estimation to demonstrate the potential of this dataset;
(3) a 3D reconstruction module based on our predictions for single-view 3D scene recovery;
(4) a generalization performance of our trained model and investigate its applicability in real-world photogrammetric scenarios.
ArchDepth is a large-scale dataset of synthetic RGB images and corresponding high-quality and complete metric depth maps. It contains ca 24.000 images of photorealistic outdoor scenes of historic buildings. The dataset is divided into two main parts:
1. Churches - six photorealistic outdoor scenes of historical churches and
2. Piazza - four scenes of an imaginary plaza composed of several buildings facades, trees, walls and statues fused into a single fictional environment.
All the models were retrieved and are freely downloadable from the Sketchfab.com 3D models sharing platform.
Download ~18.23 GB
DOWNLOADS - Samples of the Churches subset (RGB+depthmaps and 3D models)
The Churches subset of the ArchDepth dataset is comprised of a sequence of RGB images and correspondent depth maps generated in Blender. The 3D models used in the data set were retrieved from Sketchfab, created by the 3D artist Lassi Kaukonen: Kuusisto, Liedon, Mietoinen, Nousiainen, Piikkio, Saint Jacobs
This Churches subset features 15.000 RGB/depth couples subdivided per church and per rendering camera.
Kuusisto
cam1 - 640x480 pinhole camera 36mm sensor size, 24mm focal length: Download ~430 MB
cam2 - 640x480 pinhole camera 36mm sensor size, 18mm focal length: Download ~385 MB
cam3 - 640x480 pinhole camera 36mm sensor size, 14mm focal length: Download ~385 MB
cam4 - 640x480 pinhole camera 36mm sensor size, 24mm focal length: Download ~388 MB
Liedon
cam1 - 640x480 pinhole camera, 36mm sensor size, 24mm focal length: Download ~469 MB
cam2 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~417 MB
cam3 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~458 MB
cam4 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~428 MB
cam5 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length Download ~216 MB
cam5var - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~215 MB
Mietoinen
cam1 - 640x480 pinhole camera, 36mm sensor size, 24mm focal length: Download ~440 MB
cam2 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~453 MB
cam3 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~432 MB
cam4 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~460 MB
Nousiainen
cam1 - 640x480 pinhole camera, 36mm sensor size, 24mm focal length: Download ~532 MB
cam2 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~453 MB
cam3 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~500 MB
cam4 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~462 MB
Piikkio
cam1 - 640x480 pinhole camera, 36mm sensor size, 24mm focal length: Download ~455 MB
cam2 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~421 MB
cam3 - 640x480 pinhole camera, 36mm sensor size, 14mm focal length: Download ~392 MB
cam4 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~432 MB
SaintJacobs
cam1 - 640x480 pinhole camera, 36mm sensor size, 24mm focal length: Download ~376 MB
cam2 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~413 MB
cam3 - 640x480 pinhole camera, 36mm sensor size, 14mm focal length: Download ~400 MB
cam4 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~426 MB
The Churches 3D models can be downloaded here: download ~11 GB or as separate zip files:
DOWNLOAD - Samples of the Piazza subset (RGB+depthmaps and 3D models)
This Piazza subset part of the ArchDepth dataset is comprised of a sequence of RGB images and correspendent depth maps generated in Blender. The 3D models used in the subdata were retrieved from Sketchfab, created by various 3D artists: Cathédrale Saint-Pierre, Poitiers, Entry to Church of Saints Cyril and Methodius, Eglise notre-dame-la-grande, Poitiers, Eglise Sainte Radegonde, Poitiers, Eglise du Puch, Sauveterre de Guyenne, Portada románica de la catedral de Valencia, Portail roman, Saint-André-de-Sorède, Eglise Saint Sauveur, Figeac, Eglise St-Médard, Thouars, Damaged Wall, Stone Wall Nr.2, Old Tree, Tree, oak trees, Löwe, Löwe.
The Piazza subset features 9.000 RGB/depth couples subdivided per cardinal directions and per rendering cam.
CityE
cam3 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~723 MB
cam4 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~947 MB
CityN
cam1 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~718 MB
cam2 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~924 MB
CityS
cam7 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~839 MB
cam8 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~1 GB
cam9 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~793 MB
cam10 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~960 MB
CityW
cam5 - 640x480 pinhole camera, 36mm sensor size, 18mm focal length: Download ~692 MB
cam6 - 640x480 pinhole camera, 36mm sensor size, 20mm focal length: Download ~907 MB
The Piazza 3D models can be downloaded here: download ~8.5 GB or as separate zip files:
Representative Publications:
Welponer, M., Stathopoulou, E.K., Remondino, F., 2022: MONOCULAR DEPTH PREDICTION IN PHOTOGRAMMETRIC APPLICATIONS. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2022, 469–476, https://doi.org/10.5194/isprs-archives-XLIII-B2-2022-469-2022 - ISPRS Congress 2022
Funding: Internal
