In this project, we aim to learn a neural network that takes images as input and outputs a corresponding mesh.
we use a backbone network based on SRT[1].
For more deatails, please refer to the slides. You can also download a single mesh overfitting result here and a generalization result here.
Each run's config, checkpoints, and visualization are stored in a dedicated directory. Recommended configs can be found under runs/[dataset]/[model].
To train a model on a single GPU, simply run e.g.:
python3 -m torch.distributed.launch \
--nproc_per_node=$N_GPUS --master_port=$port train_sdf.py $config
To evaluate a model:
python3 -m torch.distributed.launch \
--nproc_per_node=$N_GPUS --master_port=$port eval_sdf.py $config
[1]Scene Representation Transformer: Geometry-Free Novel View Synthesis Through Set-Latent Scene Representations