Home

Awesome

3D-RETR: End-to-End Single and Multi-View 3D Reconstruction with Transformers (BMVC 2021)

Zai Shi*, Zhao Meng*, Yiran Xing, Yunpu Ma, Roger Wattenhofer

∗The first two authors contribute equally to this work

[BMVC (with presentation)] [Paper] [Supplementary]

image

Citation

@inproceedings{3d-retr,
  author    = {Zai Shi, Zhao Meng, Yiran Xing, Yunpu Ma, Roger Wattenhofer},
  title     = {3D-RETR: End-to-End Single and Multi-View3D Reconstruction with Transformers},
  booktitle = {BMVC},
  year      = {2021}
}

Create Environment

git clone git@github.com:FomalhautB/3D-RETR.git
cd 3D-RETR
conda env create -f config/environment.yaml
conda activate 3d-retr

Prepare Data

ShapeNet

Download the Rendered Images and Voxelization (32) and decompress them into $SHAPENET_IMAGE and $SHAPENET_VOXEL

Train

Here is an example of reproducing the result of the single view 3D-RETR-B on the ShapeNet dataset:

python train.py \
    --model image2voxel \
    --transformer_config config/3d-retr-b.yaml \
    --annot_path data/ShapeNet.json \
    --model_path $SHAPENET_VOX \
    --image_path $SHAPENET_IMAGES \
    --gpus 1 \
    --precision 16 \
    --deterministic \
    --train_batch_size 16 \
    --val_batch_size 16 \
    --num_workers 4 \
    --check_val_every_n_epoch 1 \
    --accumulate_grad_batches 1 \
    --view_num 1 \
    --sample_batch_num 0 \
    --loss_type dice \