Views
No views yet
News: The EMVSNet paper has been accepted for the ISPRS 2026 Toronto Congress.
train.py, models/)eval.py, evaluation/)fusion.py, fusion_padding.py)
1conda create -n emvsnet python=3.9 -y
2conda activate emvsnet
3pip install --upgrade pip
4pip install -r requirements.txt$TRAINPATH/<scan folders> from the DTU preprocessing used by this repo):export TRAINPATH=/path/to/mvs_training/dtuenv.sh and reuse shell scripts from scripts/.1python train.py \
2 --dataset dtu_yao \
3 --trainpath "$TRAINPATH" \
4 --trainlist lists/dtu/train.txt \
5 --vallist lists/dtu/val.txt \
6 --testlist lists/dtu/test.txt \
7 --batch_size 1 \
8 --epochs 1 \
9 --view_num 5 \
10 --numdepth 128 \
11 --interval_scale 1.06 \
12 --image_scale 0.25 \
13 --lr 0.001 \
14 --optimizer adam \
15 --evidential_method der \
16 --logdir ./checkpoints/quickstart1python eval.py \
2 --dataset data_eval_transform \
3 --testpath "$TRAINPATH" \
4 --testlist lists/dtu/test.txt \
5 --batch_size 1 \
6 --numdepth 256 \
7 --interval_scale 1.0 \
8 --image_scale 1.0 \
9 --view_num 7 \
10 --loadckpt /path/to/model.ckpt \
11 --outdir ./outputs_dtu \
12 --evidential_method der1python fusion.py \
2 --testpath "$TRAINPATH" \
3 --testlist lists/dtu/test.txt \
4 --outdir ./outputs_dtu \
5 --test_dataset dtudocker build -t emvsnet:latest ./workspace/data, with DTU at /workspace/data/mvs_training/dtu.1docker run --gpus all --rm -it \
2 -v /path/to/host_data:/workspace/data:ro \
3 -v /path/to/host_output:/workspace/output \
4 emvsnet:latest \
5 python train.py \
6 --dataset dtu_yao \
7 --trainpath /workspace/data/mvs_training/dtu \
8 --trainlist lists/dtu/train.txt \
9 --vallist lists/dtu/val.txt \
10 --testlist lists/dtu/test.txt \
11 --batch_size 1 \
12 --epochs 1 \
13 --view_num 5 \
14 --numdepth 128 \
15 --interval_scale 1.06 \
16 --image_scale 0.25 \
17 --lr 0.001 \
18 --optimizer adam \
19 --evidential_method der \
20 --logdir /workspace/output/checkpoints/quickstart1cp docker.env.example docker.env
2# edit docker.env (HOST_DATA_PATH and OUTPUT_PATH)
3docker compose up -d
4docker compose logs -fscripts/train_dtu.shscripts/train_dtu_ddp.shscripts/eval_dtu.shscripts/fusion_dtu.shscripts/eval_tnt.shscripts/fusion_tnt.sh1@inproceedings{emvsnet2026,
2 author = {Grannemann, Christian and Mehltretter, Max},
3 journal = {The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences},
4 title = {EMVSNet: Evidential Multi-View Stereo Reconstruction for Sampling-free Depth and Uncertainty Estimation},
5 year = {2026},
6 note = {Accepted for publication.}
7}LICENSE.