LibrePAGEsp-gazetarget

PaGE gaze-target estimation (DINOv3 ViT-S+ towers, 73M parameters), repackaged for LibreYOLO. Predicts, per person, a 64x64 heatmap of where in the scene that person is looking plus an in/out-of-frame probability.

Source

Derived from OctopusWen/PaGE at commit 4352bead, sourced from the upstream HuggingFace release Octopus1/page-vitsplus (paper: arXiv:2607.04860). Copyright (c) 2026 Zhoutong Ye / PaGE Authors. The PaGE decoder is licensed under the MIT License.

The tower parameters are derivatives of facebookresearch/dinov3 (produced by full-parameter training of DINOv3) and remain governed by the DINOv3 License: commercial use is permitted, but any redistribution must include the license text, and the weights cannot be relicensed as plain MIT.

Modifications

State-dict key normalization only (the transformers-5.x nested DINOv3 layer naming is flattened to LibreYOLO's canonical form). Learned parameters are unchanged. See weights/convert_page_weights.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibrePAGEsp-gazetarget.pt")
r = model("scene.jpg")                                  # auto head detection
r = model("scene.jpg", head_boxes=[(x1, y1, x2, y2)])   # BYO head boxes
print(r.gazetarget.xy, r.gazetarget.inout)              # targets + in-frame prob

License

MIT (decoder) + DINOv3 License (towers). See LICENSE, DINOv3_LICENSE.md and NOTICE.