CasMVSNet_pl Details

Author(s):

Quei-An Chen

Description:

A reimplementation of CasMVSNet (https://arxiv.org/pdf/1912.06378.pdf) using pytorch-lightning.

https://github.com/kwea123/CasMVSNet_pl

Method parameters:

Pretriained on DTU dataset only.

Training parameters:
image size: (640, 512)
num_views: 5
depth_interval: 2.00
depth_interval_ratio: 1.0 2.0 4.0
num_depths: 8 32 48
num_groups: 1

Testing parameters (on this dataset):
image size: (1920 or 2048, 1056)
num_views: 5
depth_interval: pre-calculated by MVSNet: can be downloaded https://drive.google.com/file/d/1YArOJaX9WVLJh4757uE8AEREYkgszrCo/view
depth_interval_ratio: 1.0 2.0 4.0
num_depths: 8 32 48
num_groups: 1
(for point cloud fusion)
confidence: 0.999
min_consistent_views: 5

runtime per image: depth prediction 1.3s, fusion 1.5s

All parameters settings can be found in my github

Side-notes:

Please follow and star if you use the code

Used image set:

Reconstructions from provided image set

Publication:

None

Software or method website:

https://github.com/kwea123/CasMVSNet_pl

---