TSDF pytorch

Simple implementation of TSDF（Truncated Signed Distance Function）in pytorch

DataSets

MVS images and cameras comes from here. Data is preprocessed by MVSNet. Manual masks are from IDR. Depth results used here are from PatchMatchNet. The normalized parameter norm_param.pkl is used to ensure that the target mesh is displayed in a -1 to 1 cube. Demo data of scan114 can be downloaded in here.

You can also get results of real-world data with cameras and depths.

DTU images (not used in fusion)

DTU depths

gt depth	patchmatchnet depth

Fusion with pytorch

The volumetric fusion costs about 2s for each view in GPU.

Fusion with predicted depth:

CUDA_VISIBLE_DEVICES=0 python run.py --data_dir <your DTU path> --pred_depth_dir <your prediction depth>

Fusion with ground truth depth:

CUDA_VISIBLE_DEVICES=0 python run.py --use_gt_depth --data_dir <your DTU path>

Fusion results of PatchMatch predicted depth