render error

[zhouilu]

i train model, dataset like mipnerf-360. train config is

CUDA_VISIBLE_DEVICES=$1 python main.py experiment/dataset=llff_360 \
    experiment/training=donerf_tensorf \
    experiment.training.val_every=5 \
    experiment.training.render_every=10 \
    experiment.training.test_every=10 \
    experiment.training.ckpt_every=80 \
    experiment/model=donerf_sphere \
    experiment.params.print_loss=True \
    experiment.dataset.collection=$3 \
    experiment.params.data_dir=$2 \
    experiment.dataset.data_subdir=data \
    +experiment/regularizers/tensorf=tv_4000

then render, config is

CUDA_VISIBLE_DEVICES=$1 python main.py experiment/dataset=llff_360 \
    experiment/training=llff_tensorf \
    experiment.training.val_every=1 \
    experiment.training.render_every=1 \
    experiment.training.test_every=1 \
    experiment/model=donerf_sphere \
    experiment.params.print_loss=True \
    experiment.dataset.collection=$3 \
    experiment.params.data_dir=$2 \
    experiment.dataset.data_subdir=data \
    +experiment/regularizers/tensorf=tv_4000 \
    experiment.params.render_only=True \
    experiment.params.save_results=True \
    experiment.training.num_epochs=80 \
    experiment.training.num_iters=80

but result is confusing. logs/val_images is good, logs/val_videos is bad result. What may cause this?

[benattal]

Hi! The results on the llff_360 dataset are likely not very good. In general, our method struggles with scenes that have a sparser distribution of cameras (e.g. inward-facing 360), due to the fact that ray-dependent sample network needs dense supervision in order to interpolate effectively. See the limitations section on page 8 of our paper for more details.