Method param_transform

gsplat/compression/yuv_codec.py:102–199

Transform the splats parameters to new domain e.g. for SH coefficients, transform them from RGB domain to YCbCr domain for quats: 1) normalize to unit quaternion 2) transform to Euler angles or other equivalent representations (Optional, depends on k

(self, splats_list: List[Dict], transform_attributes: Dict[str, bool])

        #self.compress_dir.mkdir(parents=True, exist_ok=True)
    
    def param_transform(self, splats_list: List[Dict], transform_attributes: Dict[str, bool]) -> List[Dict]:
        """Transform the splats parameters to new domain
        e.g.  
        for SH coefficients, transform them from RGB domain to YCbCr domain
        for quats:
            1) normalize to unit quaternion
            2) transform to Euler angles or other equivalent representations (Optional, depends on kwargs)
    
        Args:                                                              
            splats_list (List[Dict]): splats parameters of all frames
            transform_attributes (Dict[str, bool]): indicate attributes whether to transform domain
        """
        transformed_splats_list = []
        
        # Define the RGB to YCbCr transformation matrices for different standards
        
        # BT.709 (Rec. 709) standard, which is the most widely used for HD content
        # Reference: ITU-R BT.709-6 (06/2015)
        # 
        # Key characteristics:
        # - Designed for HD and Full HD content (1080p)
        # - Uses D65 white point
        # - Assumes RGB values are in range [0,1]
        # - Output Y in range [0,1], Cb/Cr in range [-0.5,0.5]
        bt709_rgb_to_ycbcr = torch.tensor([
            [ 0.299,    0.587,    0.114  ],  # Y  = 0.299R  + 0.587G  + 0.114B
            [-0.169,   -0.331,    0.5    ],  # Cb = -0.169R - 0.331G  + 0.5B
            [ 0.5,     -0.419,   -0.081  ]   # Cr = 0.5R    - 0.419G  - 0.081B
        ], dtype=torch.float32)
        
        # BT.470 (older standard for SD/analog TV content)
        # Reference: ITU-R BT.470
        # 
        # Key characteristics:
        # - Designed for SD content and analog TV broadcasts
        # - Uses different chroma scaling compared to BT.709
        # - Assumes RGB values are in range [0,1]
        # - Output Y in range [0,1], Cb/Cr in range [-0.5,0.5]
        bt470_rgb_to_ycbcr = torch.tensor([
            [ 0.299,    0.587,    0.114  ],  # Y  = 0.299R  + 0.587G  + 0.114B
            [-0.147,   -0.289,    0.436  ],  # Cb = -0.147R - 0.289G + 0.436B
            [ 0.615,   -0.515,   -0.100  ]   # Cr = 0.615R - 0.515G - 0.100B
        ], dtype=torch.float32)
        
        # Select the appropriate transformation matrix based on color standard
        if self.color_standard == "BT470":
            rgb_to_ycbcr = bt470_rgb_to_ycbcr
            if self.verbose:
                print("Using BT470 RGB to YCbCr transformation matrix")
        else:  # Default to BT709
            rgb_to_ycbcr = bt709_rgb_to_ycbcr
            if self.verbose:
                print("Using BT709 RGB to YCbCr transformation matrix")
        
        for splats in splats_list:
            transformed_splats = {}
            
            # Process all keys in the splats dictionary