Function render

gaussian_renderer_dynamic/__init__.py:64–193

(gaussians: dict, bg_color: torch.Tensor, timestamps: torch.Tensor = None, scaling_modifier=1.0, 
           opt: Options=None, anchor_time: torch.Tensor=None,
           training=True, override_opacity=False,
           )

tanfovy = math.tan(math.pi / 4.0)

def render(gaussians: dict, bg_color: torch.Tensor, timestamps: torch.Tensor = None, scaling_modifier=1.0, 
           opt: Options=None, anchor_time: torch.Tensor=None,
           training=True, override_opacity=False,
           ):
    
    # random background color augmentation
    if training:
        bg_color = torch.rand(3).cuda()
    else:
        # bg_color = torch.tensor([1.0, 1.0, 1.0]).cuda()
        bg_color = torch.tensor([0.5, 0.5, 0.5]).cuda()

    # bg_color = torch.tensor([0.5, 0.5, 0.5]).cuda() 
    L = 0
    LP = opt.forder
    
    batch_size, gaussian_num = gaussians['xyz'].shape[0], gaussians['xyz'].shape[1]
    
    screenspace_points = torch.zeros_like(gaussians['xyz'][:, :, 0, :], dtype=gaussians['xyz'].dtype, requires_grad=True, device=gaussians['xyz'].device)
    screenspace_points.retain_grad()

    view_matrix = world_view_transform.float()  # View matrix
    view_proj_matrix = full_proj_transform.float()  # Projection matrix
    campos = camera_center.float()  # Camera position

    if len(opt.down_resolution) > 0:
        render_height, render_width = opt.down_resolution
    else:
        render_height, render_width = opt.image_height, opt.image_width
    raster_settings = GaussianRasterizationSettingsOrth(
        image_height=render_height,
        image_width=render_width,
        tanfovx=tanfovx,
        tanfovy=tanfovy,
        bg=bg_color if bg_color is not None else bg_color,
        scale_modifier=scaling_modifier,
        viewmatrix=view_matrix.cuda(),
        projmatrix=view_proj_matrix.cuda(),
        sh_degree=0,  
        campos=campos.cuda(),
        prefiltered=False,
        debug=False,
    )
    
    rasterizer = GaussianRasterizerOrth(raster_settings=raster_settings)
    render_images = []
    render_depths = []
    render_alphas = []

    dummy_time = torch.zeros(1, device=gaussians['xyz'].device)
    output_frames = opt.output_frames
    N = gaussians['xyz'].shape[1]
    if timestamps is None:
        output_frames = 1
        timestamps = dummy_time.repeat(batch_size, output_frames)
    if anchor_time is None:
        anchor_time = torch.zeros((N, 1), device=gaussians['xyz'].device)
    else: