Quantitative results of reconstruction quality on the DTU dataset. in terms of the distance metric(lower is better) and the percentage metric (see Descriptions for more details) with 1mm and 2mm as thresholds. To understand the experiment settings and evaluation rules, check Submissions and Description.

To show the leaderboard at each Sparsity, just search the sparsity index (e.g., search "Five" if sparsity=5 in the search bar and then click the matric you want to rank.). To show the performance changes with sparsity for each method,just click the method colume. The reference and the project page of each method can be find in Reference.

Ranking Method Mean Distance(mm) Percentage(<1mm) Percentage(<2mm)
Precision Recall Overall Precision Recall f-score Precision Recall f-score

You can download the full table in csv format. Download the table as csv

Chart Comparision

The following chart plots the performance in terms of f-score (1mm), which unifies both recall and precision, which corresponds to the results in the tables above.

You can click the method in the legend to hide/show each of them.

F_score comparison

Comparison with the existing methods in the DTU dataset.


Method Reference Project Link
SurfaceNet+ SurfaceNet+: An End-to-end 3D Neural Network for Very Sparse Multi-view Stereopsis SurfaceNet-plus
SurfaceNet SurfaceNet: An End-to-end 3D Neural Network for Multiview Stereopsis SurfaceNet
Gipuma Massively Parallel Multiview Stereopsis by Surface Normal Diffusion Gipuma
COLMAP Pixelwise View Selection for Unstructured Multi-View Stereo (2016 CVPR) COLMAP
R-MVSNet Recurrent MVSNet for High-resolution Multi-view Stereo Depth Inference R-MVSNet