haitian2du / paver360sod

Official repository of Panoramic Vision Transformer for Saliency Detection in 360° Videos (ECCV 2022)

PAVER

Official repository of Panoramic Vision Transformer for Saliency Detection in 360 Videos (ECCV 2022)

[Paper] [Poster]

Getting Started

1. To create a virtual environment with all relevant libraries installed:

conda env create -f environment.yml

2. Dataset will be stored under the data directory by default. First, create a symbolic link at your convenience:

ln -s /some/path/to/data data

3. The data directory should be organized as follows. Since the Wild360 dataset is available upon request, you may ask for access here.

data
├─ wild360                # the dataset should be downloaded on your own
│  ├─ data/{test,train}
│  └─ Wild360_GT_29
├─ cache                  # cache features for fast experiments
└─ log                    # training log & weights

How to Run

Training

Provided the dataset and environment are all prepared, simply run the command below:

python cli.py

Default configurations can be found in config.py and configs/*/*.json. To override parameters at your disposal:

python cli.py with max_epoch=10

Inference

You can download our model weight here.

python cli.py demo with {some_necessary_parameters}

Adapting PAVER for ViT Variants

Adapting PAVER for ViT variants is quite simple. Provided the target ViT uses PatchEmbed, all you need to do is to replace it with a deformable counterpart using compute_deform_offset in geometry.py.

# (...)
from geometry import compute_deform_offset

class DeformPatchEmbed(nn.Module):
    def __init__(self, ...):
        # (...)
        self.proj = DeformConv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        offset = torch.from_numpy(compute_deform_offset(model_config=model_config,
                                                        is_discrete=is_discrete)).float()
        self.register_buffer('offset', offset)
        # (...)

class DeformViTVariant(ViT):
    def __init__(self, ...):
        # (...)
        self.patch_embed = DeformPatchEmbed(...)
        # (...)

Please refer to model/ViT.py for detailed usage.

Citation

If you find our work useful in your research, please consider citing:

@InProceedings{
    author    = {Yun, Heeseung and Lee, Sehun and Kim, Gunhee},
    title     = {Panoramic Vision Transformer for Saliency Detection in 360 Videos},
    booktitle = {ECCV},
    year      = {2022}
}