What it is
Self-supervised ViT backbones pretrained for dense spatial perception, exposing patch features that preserve boundaries, shapes, and semantic regions. Provides pretrained ViT visual encoders for extracting dense patch tokens that align with object boundaries and spatial structure, with loaders and PCA demos for downstream depth, segmentation, and token matching.
Key results
- Released backbones span ViT-S/16 to 1.1B-parameter ViT-g/16
- Extracts patch tokens as [B, H*W, C]; example: [1, 1024, 384] at 512 px
- Backbone-only .pt checkpoints auto-download from Hugging Face
Shows self-supervised ViT pretraining can produce frozen features that capture geometry and semantics together for dense perception.