ReVA-HPE: Relative 6DoF Head Pose Estimation with Synthetic Pose Supervision and Personalized Virtual Anchors

WACV 2027 · Paper · Supplementary · Project page · Code

Vasiliki Vasileiou, Panagiotis P. Filntisis, Petros Maragos, Kostas Daniilidis

ReVA-HPE estimates metric 6DoF head pose without a ground-truth anchor pose. From a single unlabeled enrollment image we reconstruct a personalized Gaussian avatar and render it into a bank of virtual anchors with known poses: query rotation comes from pose-consensus RANSAC over the bank, and metric translation from one enrollment landmark-PnP solve combined with the learned depth-normalized relative translation. The relative estimator — the frozen VGGT backbone adapted with LoRA, plus a fine-tuned camera head — is trained purely on procedural FLAME renderings. A fast variant (VA-fast) uses the bank once at enrollment and then needs a single real-to-real prediction per query.

Files

File	Size
checkpoint_410.pt — the relative 6DoF estimator (the paper model)	6.3 GB

The checkpoint contains VGGT-1B-derived weights and is released under the VGGT research (non-commercial) license.

Usage

The checkpoint is consumed by the evaluation and training code in the GitHub repository:

git clone https://github.com/VasilikiVas/ReVA-HPE
cd ReVA-HPE
conda env create -f environment.yml && conda activate reva-hpe
cp assets.env.example assets.env   # point REVA_HPE_CHECKPOINT at checkpoint_410.pt

# BIWI, practical ReVA-HPE (Direct-5r):
scripts/eval_biwi_direct5r_subject.sh 1      # per subject, 1..24
scripts/aggregate_biwi_direct5r.sh
scripts/eval_biwi_pnp_relative_subject.sh 1
scripts/fuse_biwi_reva.sh

You also need the frozen VGGT-1B base checkpoint from facebookresearch/vggt; datasets and the avatar-bank pipeline are documented in the repository's DATA.md and CHECKPOINTS.md.

Results (shared MTCNN protocol)

Rotation MAE in degrees, translation MAE in millimeters. Best results in bold. "First frame" / "near-frontal" refer to the enrollment selection.

Method	BIWI MAEr	BIWI MAEt	FaMoS MAEr	Pose supervision
6DRepNet	3.91	—	5.98	mixed
SemiUHPE	3.20	—	2.86	real
TRG	5.15	48.71	5.59	mixed
ReVA-HPE (first frame)	2.53	17.10	3.78	synthetic
ReVA-HPE (near-frontal)	2.49	18.94	2.47	synthetic

Citation

@inproceedings{vasileiou2027revahpe,
  title={ReVA-HPE: Relative 6DoF Head Pose Estimation with Synthetic Pose Supervision and Personalized Virtual Anchors},
  author={Vasileiou, Vasiliki and Filntisis, Panagiotis P. and Maragos, Petros and Daniilidis, Kostas},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year={2027}
}