Dense Correspondence Adaptation Artifacts

This repository stores the external artifacts for the Dense Correspondence Adaptation paper release. The source code lives in the companion GitHub repository and is released under Apache-2.0.

Contents

  • models/<model_tag>/merged_backbone.pt: merged timm-compatible backbone weights for trained models.
  • metrics/<model_tag>/*.json: TAP, DAVIS VOS, and Objaverse metrics artifacts used for paper tables.
  • paper_manifest.json: machine-readable mapping from model tags to configs, commands, expected metrics, files, and checksums.
  • sha256sums.txt: checksum list for uploaded files.

The DINOv2 pretrained baseline uses timm/vit_base_patch14_dinov2.lvd142m at revision 4685c99dabffe5affac90bd99dbffd25801ae58d and therefore has no uploaded checkpoint. Objaverse metrics generated with --include-table-objaverse-metrics are marked with provenance.kind = paper_table_reconstruction; they can be regenerated with the public Objaverse evaluator from documented local data paths.

Paper Metrics

Model tag Config TAP AJ TAP APT DAVIS J&F Objaverse PCDP@0.05
dinov2_pretrained_baseline configs/paper/dinov2_pretrained_baseline.yaml 0.2009 0.3045 0.7030 35.90
dinov2_davis_split_head_corr005 configs/paper/dinov2_davis_split_head_corr005.yaml 0.2518 0.3730 0.6397 31.80
dinov2_davis_projection_head_corr005 configs/paper/dinov2_davis_projection_head_corr005.yaml 0.2502 0.3711 0.6372 30.76

Use

Install the public package, download the relevant merged_backbone.pt, and run the documented CLI commands from paper_manifest.json. Users must obtain datasets from their original sources and comply with dataset terms.

License And Terms

Source code and released trained backbones use Apache-2.0. Dataset videos, TAP-Vid annotations, DAVIS masks, YouTube-VOS data, Objaverse renderings, and third-party checkpoints are not redistributed here.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support