This directory contains a raw data sample. Not in 300W-LP format.
Labels are generated by export_face_params in replicantface/export.py
np.savez_compressed(str(label_path),
modelview = np.asarray(m_cam, dtype=np.float32),
projection = np.asarray(proj_matrix, dtype=np.float32),
vertices = np.asarray(vertices, dtype=np.float16),
resolution = render.resolution_x)The vertices are defined w.r.t. the models local coordinates. The modelview matrix (4x4) describes the position and
orientation of the head relative to the camera. projection is the usual projection matrix (4x4) for pinhole cameras
from computer graphics, mapping to [-1,1] normalized screen space. resolution is just the image size, currently always
640, added for convenience.
The segmentation colors should have been primary colors, but due to how Blender apparently works and me not noticing early enough, they ended up distorted. At boundaries, they are blended smoothly.
COLOR_FACE = (220,57,33)
COLOR_BEARD = (118,190,70)
COLOR_CLOTHES = (135,198,199)
COLOR_BG = (0,0,0)face_{counter:5d}.npzas described above
face_{counter:5d}_img.npzthe visual images
face_{counter:5d}_mask.npzthe visual images

