BlendedMVS: A Large-Scale Dataset for Generalized Multi-View Stereo Networks

By Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, Long Quan. Computer Vision and Pattern Recognition, 2019

756 citations recorded by the source corpus.

While deep learning has recently achieved great success on multi-view stereo (MVS), limited training data makes the trained model hard to be generalized to unseen scenarios. Compared with other computer vision tasks, it is rather difficult to collect a large-scale MVS dataset as it requires expensive active scanners and labor-intensive process to obtain ground truth 3D structures. In this paper, we introduce BlendedMVS, a novel large-scale dataset, to provide sufficient training ground truth for learning-based MVS. To create the dataset, we apply a 3D reconstruction pipeline to recover high-quality textured meshes from images of well-selected scenes. Then, we render these mesh models to color images and depth maps. To introduce the ambient lighting information during training, the rendered color images are further blended with the input images to generate the training input. Our dataset contains over 17k high-resolution images covering a variety of scenes, including cities, architectures, sculptures and small objects. Extensive experiments demonstrate that BlendedMVS endows the trained model with significantly better generalization ability compared with other MVS datasets. The dataset and pretrained models are available at https://github.com/YoYo000/BlendedMVS.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/e32541301130923df1ae4a83cb8037bd459e5254

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass (2025, positive): “Fast3R’s predictions of ( ˆ X L , ˆΣ L , ˆ X G , ˆΣ G ) are trained using generalized versions of the pointmap loss in DUST3R [61].”
Vision Transformers for Dense Prediction (2021, methodology): “Similar to [32], we first pretrain on a well-curated subset of the data [47, 48, 49] for 60 epochs before training on the full dataset.”
Neural Sparse Voxel Fields (2020, methodology): “Details on the original source ﬁles and license information are given below: – • BlendedMVS.”
DL3DV-10K: A Large-Scale Scene Dataset for Deep Learning-based 3D Vision (2023, neutral): “Scene-level datasets are essential for NVS and scene understanding, yet offerings like LLFF [18], Matter-port3D [6], and BlendedMVS [34] encompass limited scenes.”
ZoeDepth: Zero-shot Transfer by Combining Relative and Metric Depth (2023, methodology): “…the relative depth backbone, we train on a mix of 12 datasets (M12) consisting of the 10 datasets used in [32]: HRWSI [46], Blend-edMVS [47], ReDWeb [45], DIML-Indoor [17], 3D Movies [33], MegaDepth [25], WSVD [41], TartanAir [43], ApolloScape [15] and IRS [42], plus 2 additional…”
Unified-IO: A Unified Model for Vision, Language, and Multi-Modal Tasks (2022, background): “, 2019a) and BlendedMVS (Yao et al., 2020) surface normal estimation datasets.”
KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs (2021, methodology): “Consequently, KiloNeRF is tested on both synthetic datasets (Synthetic-NeRF, Synthetic-NSVF) as well as datasets comprising real scenes (NSVF’s variants of BlendedMVS [71] and Tanks&Temples [18]).”
VGGT: Visual Geometry Grounded Transformer (2025, neutral): “Hence, while the feed-forward mode of VGGT outperforms all previous alternatives (whether they are feed-forward or not), there is still room for improvement since post-optimization still brings beneﬁts.”
Omnidata: A Scalable Pipeline for Making Multi-Task Mid-Level Vision Datasets from 3D Scans (2021, background): “…datasets: Indoor scene datasets: Replica [56], HyperSim [49], Taskonomy [72], Habitat-Matterport (HM3D) Aerial/outdoor datasets: BlendedMVG [69] Diagnostic/Structured datasets: CLEVR [28] Object-centric datsets: To provide object-centric views in addition to scene-centric ones, we create a…”
Depth Anything 3: Recovering the Visual Space from Any Views (2025, neutral): “AriaDigitalTwin [63] 237 Synthetic AriaSyntheticENV [63] 99950 Synthetic ArkitScenes [5] 4388 LiDAR BlendedMVS [115] 503 3D Recon Co3dv2 [68] 30616 Colmap DL3DV [53] 6379 Colmap HyperSim [69] 344 Synthetic MapFree [3] 921 Colmap MegaDepth [51] 268 Colmap MegaSynth [40] 6049 Synthetic MvsSynth [38]…”
Depth Anything V2 (2024, neutral): “Moreover, BlendedMVS [92] is critical to the capability of dealing with the bird’s-eye view.”
NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction (2021, methodology): “We further tested on 7 challenging scenes from the low-res set of the BlendedMVS dataset [45](CC-4 License).”
Grounding Image Matching in 3D with MASt3R (2024, methodology): “We train our network with a mixture of 14 datasets: Habitat [74], ARKitScenes [20], Blended MVS [112], MegaDepth [48], Static Scenes 3D [57], ScanNet++ [113], CO3D-v2 [67], Waymo [83], Map-free [5], WildRgb [2], VirtualKitti [12], Unreal4K [91], TartanAir [103] and an internal dataset.”
Geo-Neus: Geometry-Consistent Neural Implicit Surfaces Learning for Multi-view Reconstruction (2022, methodology): “We also test on 7 challenging scenes from the low-res set of the BlendedMVS dataset [38] (CC-4 License).”
ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes (2023, methodology): “Meanwhile, datasets that were originally proposed for multi-view stereo such as DTU [15], BlendedMVS [47], and Tanks and Temples [20] are now also used for novel view synthesis.”
DUSt3R: Geometric 3D Vision Made Easy (2023, neutral): “DUSt3R is trained with a mixture of eight datasets: Habitat [104], ARKitScenes [25], MegaDepth [56], Static Scenes 3D [110], Blended MVS [162], Scan-Net++ [166], CO3Dv2 [94] and Waymo [122].”