| """Depth Anything 3 preprocessing: the package's ``InputProcessor`` (vendored in ``da3_input_processor.py``) behind the |
| transformers image-processor interface. Images (PIL / numpy / paths) -> ``pixel_values`` (1, N, 3, H, W), resized |
| ("upper_bound_resize" to 504 by default, sides multiples of 14) and ImageNet-normalised; camera parameters given |
| with the images come back resized to match (``extrinsics`` / ``intrinsics``).""" |
|
|
| from transformers.feature_extraction_utils import BatchFeature |
| from transformers.image_processing_utils import BaseImageProcessor |
|
|
| from .da3_input_processor import InputProcessor |
|
|
|
|
| class DepthAnything3ImageProcessor(BaseImageProcessor): |
| model_input_names = ["pixel_values"] |
|
|
| def __init__(self, process_res: int = 504, process_res_method: str = "upper_bound_resize", **kwargs): |
| super().__init__(**kwargs) |
| self.process_res = process_res |
| self.process_res_method = process_res_method |
|
|
| def preprocess(self, images, extrinsics=None, intrinsics=None, process_res=None, process_res_method=None, |
| return_tensors=None, **kwargs) -> BatchFeature: |
| if not isinstance(images, (list, tuple)): |
| images = [images] |
| tensor, extrinsics, intrinsics = InputProcessor()( |
| list(images), |
| extrinsics.copy() if extrinsics is not None else None, |
| intrinsics.copy() if intrinsics is not None else None, |
| process_res or self.process_res, |
| process_res_method or self.process_res_method, |
| **kwargs, |
| ) |
| |
| data = {"pixel_values": tensor[None].float()} |
| if extrinsics is not None: |
| data["extrinsics"] = extrinsics[None] |
| if intrinsics is not None: |
| data["intrinsics"] = intrinsics[None] |
| return BatchFeature(data, tensor_type=return_tensors) |
|
|
|
|
| __all__ = ["DepthAnything3ImageProcessor"] |
|
|
