kornia/sold2

Pretrained weights for SOLD² (Self-supervised Occlusion-aware Line Description and Detection), used by kornia.feature.SOLD2.

SOLD² detects line segments and computes semi-dense descriptors along them using a shared encoder. Trained on wireframe and outdoor datasets. CVPR 2021.

Original repo: cvg/SOLD2

Weights

File	Training data	Contents
sold2_wireframe.safetensors	Wireframe dataset	model weights (model_state_dict), safetensors format
sold2_wireframe.pth	Wireframe dataset	the upstream training checkpoint: weights, optimizer state, epoch, config

Provenance

sold2_wireframe.pth is byte-identical to the upstream sold2_wireframe.tar (sha256 7b8b0e712f743c03a6b1a33b194926e7e8ae5467c85de55b461f54777f24606f). sold2_wireframe.safetensors holds its model_state_dict unchanged: the same keys, dtypes, shapes and bits. The optimizer state, epoch and training config are not included. The kornia maintainers converted it (converter revision 5866995a4c) and checked that kornia.feature.SOLD2 and kornia.feature.SOLD2_detector give bitwise-identical outputs from either file. The sha256 of sold2_wireframe.safetensors is 3e2adb91351b443f054c96073f70de37d7e2d3c91727aa260c75abfb848f680e.

License

MIT, Copyright (c) 2020 Rémi Pautrat, the licence of cvg/SOLD2. See LICENSE, copied from that repository.

Citation

@inproceedings{SOLD22021,
    author    = {Pautrat*, Rémi and Lin*, Juan-Ting and Larsson, Viktor
                 and Oswald, Martin R. and Pollefeys, Marc},
    title     = {{SOLD2}: Self-supervised Occlusion-aware Line Description and Detection},
    booktitle = {CVPR},
    year      = {2021}
}