def __init__(self, cfg):
super(VoxMeshHead, self).__init__()
self.setup(cfg)
# backbone
self.backbone, feat_dims = build_backbone(
cfg.MODEL.VOXEL_HEAD.RGB_BACKBONE)
# voxel head
cfg.MODEL.VOXEL_HEAD.COMPUTED_INPUT_CHANNELS = feat_dims[-1]
self.voxel_head = VoxelHead(cfg)
# mesh head
cfg.MODEL.MESH_HEAD.COMPUTED_INPUT_CHANNELS = sum(feat_dims)
self.mesh_head = MeshRefinementHead(cfg)
def __init__(self, cfg):
super(Pixel2MeshHead, self).__init__()
# fmt: off
backbone = cfg.MODEL.MESH_HEAD.RGB_BACKBONE
self.ico_sphere_level = cfg.MODEL.MESH_HEAD.ICO_SPHERE_LEVEL
# fmt: on
self.register_buffer("K", get_blender_intrinsic_matrix())
# backbone
self.backbone, feat_dims = build_backbone(backbone)
# mesh head
cfg.MODEL.MESH_HEAD.COMPUTED_INPUT_CHANNELS = sum(feat_dims)
self.mesh_head = MeshRefinementHead(cfg)
def __init__(self, cfg):
super(VoxMeshHead, self).__init__()
# fmt: off
backbone = cfg.MODEL.BACKBONE
self.cubify_threshold = cfg.MODEL.VOXEL_HEAD.CUBIFY_THRESH
self.voxel_size = cfg.MODEL.VOXEL_HEAD.VOXEL_SIZE
# fmt: on
self.register_buffer("K", get_blender_intrinsic_matrix())
# backbone
self.backbone, feat_dims = build_backbone(backbone)
# voxel head
cfg.MODEL.VOXEL_HEAD.COMPUTED_INPUT_CHANNELS = feat_dims[-1]
self.voxel_head = VoxelHead(cfg)
# mesh head
cfg.MODEL.MESH_HEAD.COMPUTED_INPUT_CHANNELS = sum(feat_dims)
self.mesh_head = MeshRefinementHead(cfg)
def __init__(self, cfg):
nn.Module.__init__(self)
self.setup(cfg)
self.single_view_voxel_prediction = cfg.MODEL.VOXEL_HEAD.SINGLE_VIEW
# backbone
self.backbone, feat_dims = build_backbone(
cfg.MODEL.VOXEL_HEAD.RGB_BACKBONE)
# voxel head
cfg.MODEL.VOXEL_HEAD.COMPUTED_INPUT_CHANNELS = feat_dims[-1]
self.voxel_head = VoxelHead(cfg)
# multi-view feature fusion
prefusion_feat_dims = sum(feat_dims)
postfusion_feat_dims = self.init_feature_fusion(
cfg, prefusion_feat_dims)
# mesh head
# times 3 cuz multi-view (mean, avg, std) features will be used
cfg.MODEL.MESH_HEAD.COMPUTED_INPUT_CHANNELS = postfusion_feat_dims
self.mesh_head = MeshRefinementHead(cfg, self.fuse_multiview_features)
def init_mesh_head(self, cfg):
self.contrastive_depth_type = cfg.MODEL.CONTRASTIVE_DEPTH_TYPE
if cfg.MODEL.MESH_HEAD.RGB_FEATURES_INPUT:
if getattr(self, "pre_voxel_rgb_cnn", None) is not None:
# reuse RGB CNN is already exists
self.post_voxel_rgb_cnn = self.pre_voxel_rgb_cnn
self.post_voxel_rgb_feat_dims = self.pre_voxel_rgb_feat_dims
else:
self.post_voxel_rgb_cnn, self.post_voxel_rgb_feat_dims \
= build_backbone(cfg.MODEL.MESH_HEAD.RGB_BACKBONE)
else:
self.post_voxel_rgb_cnn = None
self.post_voxel_rgb_feat_dims = [0]
if cfg.MODEL.MESH_HEAD.DEPTH_FEATURES_INPUT:
self.post_voxel_depth_feat_dims = self.init_post_voxel_depth_cnn(
cfg)
else:
self.post_voxel_depth_cnn = None
self.post_voxel_depth_feat_dims = [0]
# multi-view feature fusion
self.prefusion_feat_dims = sum(self.post_voxel_rgb_feat_dims) \
+ sum(self.post_voxel_depth_feat_dims)
self.postfusion_feat_dims = self.init_feature_fusion(
cfg, self.prefusion_feat_dims)
# mesh head
cfg.MODEL.MESH_HEAD.COMPUTED_INPUT_CHANNELS = self.postfusion_feat_dims
self.mesh_head = MeshRefinementHead(cfg, self.fuse_multiview_features)
# depth renderer
self.depth_renderer = DepthRenderer(cfg)
print({
"post_voxel_rgb_feat_dims": self.post_voxel_rgb_feat_dims,
"post_voxel_depth_feat_dims": self.post_voxel_depth_feat_dims,
"prefusion_feat_dims": self.prefusion_feat_dims,
"postfusion_feat_dims": self.postfusion_feat_dims
})