VGGT-HPE: Reframing Head Pose Estimation as Relative Pose Prediction

CVPRW 2026 · Paper · Project page · Code

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

VGGT-HPE reframes monocular head pose estimation as relative pose prediction: instead of regressing an absolute pose from a single image, we estimate the rigid transformation between a query image and an anchor image with known pose. The model adapts the VGGT camera branch with LoRA (rank 8) and is trained exclusively on synthetic FLAME renderings, with supervision on rotation, translation, and field of view. It achieves state-of-the-art BIWI performance, and relative prediction becomes increasingly advantageous as pose difficulty grows.

Files

File	Size
checkpoint_305.pt — the relative model (the paper model)	6.3 GB
checkpoint_abs.pt — the VGGT-HPE-Abs absolute baseline	5.87 GB

Both contain VGGT-1B-derived weights and are released under the VGGT research (non-commercial) license. Verify downloads against the SHA-256 values in the repository's CHECKPOINTS.md.

Usage

The checkpoints are consumed by the evaluation and training code in the GitHub repository:

git clone https://github.com/VasilikiVas/vggt-hpe
cd vggt-hpe
conda env create -f environment.yml && conda activate vggt-hpe
cp assets.env.example assets.env   # point VGGT_HPE_CHECKPOINT at checkpoint_305.pt

# Full BIWI (Table 1):
OUTPUT_DIR=outputs/biwi_main scripts/eval_biwi.sh

# Fixed hard / easy benchmarks (Tables 2-3):
PAIR_CSV=protocols/table2_hard_pairs.csv OUTPUT_DIR=outputs/biwi_hard scripts/eval_biwi.sh
PAIR_CSV=protocols/table3_easy_pairs.csv OUTPUT_DIR=outputs/biwi_easy scripts/eval_biwi.sh

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

Results (BIWI, shared MTCNN protocol)

Rotation MAE in degrees; best results in bold.

Method	Yaw ↓	Pitch ↓	Roll ↓	MAE ↓	Train data
6DRepNet	3.74	4.95	3.04	3.91	mixed
TokenHPE-v1	5.57	6.23	3.79	5.20	mixed
TRG	4.58	7.18	3.68	5.15	mixed
VGGT-HPE-Abs (ours)	4.90	7.01	3.53	5.15	synthetic
VGGT-HPE (Rel., ours)	2.24	3.04	3.17	2.82	synthetic

Citation

@InProceedings{Vasileiou_2026_CVPR,
    author    = {Vasileiou, Vasiliki and Filntisis, Panagiotis P and Maragos, Petros and Daniilidis, Kostas},
    title     = {VGGT-HPE: Reframing Head Pose Estimation as Relative Pose Prediction},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2026},
    pages     = {5464-5473}
}