DePoser
A pipeline that combines deep foundation models for local annotation and fine-tuning, used to adapt HMR models to end-user data distributions under privacy constraints.
Timeline (last 90 days)
DePoser is a local annotation and fine-tuning pipeline combining depth foundation models, used to adapt a collaboratively trained HMR model to the end user's own data distribution.
On in-the-wild data, fine-tuning with DePoser annotations can improve personalized accuracy.
DePoser's fine-tuning improvement is relative to the pseudo-ground truth generated by DePoser, and is not an independent validation against absolute ground truth.
In DePoser, the depth foundation model is used to generate 3D geometric supervision signals for local data, automatically producing pseudo-annotations of human meshes combined with human priors or 2D keypoints.
DePoser relies on depth foundation models for annotation, so the quality ceiling of its pseudo-ground truth is determined by the depth foundation model, and errors propagate into subsequent fine-tuning.
All Facts (5)
DePoser is a local annotation and fine-tuning pipeline combining depth foundation models, used to adapt a collaboratively trained HMR model to the end user's own data distribution.
50%UnverifiedOn in-the-wild data, fine-tuning with DePoser annotations can improve personalized accuracy.
50%UnverifiedDePoser's fine-tuning improvement is relative to the pseudo-ground truth generated by DePoser, and is not an independent validation against absolute ground truth.
50%UnverifiedIn DePoser, the depth foundation model is used to generate 3D geometric supervision signals for local data, automatically producing pseudo-annotations of human meshes combined with human priors or 2D keypoints.
50%UnverifiedDePoser relies on depth foundation models for annotation, so the quality ceiling of its pseudo-ground truth is determined by the depth foundation model, and errors propagate into subsequent fine-tuning.
50%