Skip to content

Latest commit

 

History

History
48 lines (34 loc) · 1.88 KB

INSTRUCTIONS_PIX3D.md

File metadata and controls

48 lines (34 loc) · 1.88 KB
Raw

Experiments on Pix3D

Download Pix3D and splits

Run

datasets/pix3d/download_pix3d.sh

to download Pix3D and the S1 & S2 splits to ./datasets/pix3d/

Training

python tools/train_net.py --num-gpus 8 \
--config-file configs/pix3d/meshrcnn_R50_FPN.yaml

Note that the above config is tuned for 8-gpu distributed training. Deviation from the provided training recipe means that other hyper parameters have to be tuned accordingly.

Testing and Evaluation

python tools/train_net.py \
--config-file configs/pix3d/meshrcnn_R50_FPN.yaml \
--eval-only MODEL.WEIGHTS /path/to/checkpoint_file

If you wish to evaluate the provided pretrained models (see below for a model zoo), simply do MODEL.WEIGHTS meshrcnn://meshrcnn_R50.pth. Note that by default, the config files use the S1 split.To change between S1 and S2, specify the split in the DATASETS section in the config file.

Models

We provide a model zoo for models trained on Pix3D S1 & S2 splits (see paper for more details).

Mesh R-CNN Pixel2Mesh SphereInit
S1 meshrcnn_R50.pth pixel2mesh_R50.pth sphereinit_R50.pth
S2 meshrcnn_S2_R50.pth pixel2mesh_S2_R50.pth sphereinit_S2_R50.pth