DepthAnythingv2Pipe

DepthAnythingv2Pipe is registered as depth-anything-v2-pipe and implemented in src/autopipeline/components/modules/depth_anything_v2_pipe.py.

This module computes structural similarity in depth space rather than RGB space. It is useful when geometric consistency is more important than appearance similarity.

Class

Overview

Registry Entry

Field	Value
Registry key	depth-anything-v2-pipe
Class	DepthAnythingv2Pipe
Main mixins	DepthAnythingv2Mixin, SSIMMixin, MaskProcessor
Return type	float

Constructor

Constructor

DepthAnythingv2Pipe(**kwargs)

Supported init kwargs

Key	Required	Meaning
model_path	Yes in practice	Hugging Face depth-estimation checkpoint path.
device	No	Torch device for depth inference.

Methods

Public Methods

Method	Purpose
_normalize_depth_map(depth_map)	Rescale a raw depth map into uint8 [0, 255].
calc_depth_ssim(...)	Predict depth for both images and compute masked SSIM in depth space.
__call__(...)	Dispatch to the depth_ssim branch.

Signature

Call Signature

DepthAnythingv2Pipe.__call__(
    ref_image: Image.Image,
    edited_image: Image.Image,
    coords: List[Tuple[int, int, int, int]] = None,
    mask_mode: str = None,
    metric: str = "depth_ssim",
    **kwargs,
)

Input / Output

Runtime Inputs

Argument	Required	Meaning
ref_image	Yes	Reference image.
edited_image	Yes	Edited image.
coords	No	Region boxes used for masking.
mask_mode	No	inner or outer, normally derived from pipeline scope.
metric	Yes	Currently only depth_ssim.

Extra runtime kwargs

Key	Default	Meaning
resize_depth_maps	True	Resize predicted depth back to original image size before scoring.
win_size	7	SSIM window size.
win_sigma	1.5	SSIM window sigma.

Supported Metric

Metric	What it measures	Better direction
depth_ssim	structural similarity between predicted depth maps	higher is better

Execution flow

calc_depth_ssim(...) performs the following steps:

1. predict depth maps for the reference and edited images
2. optionally resize them back to original image size
3. min-max normalize each depth map into uint8
4. build a mask with MaskProcessor
5. reduce the mask to one channel if necessary
6. compute SSIM on single-channel depth tensors

Input / Output

Return Value

The pipe returns a single float score.

As with other SSIM-backed paths:

• higher is better
• SSIMMixin.compute(...) may degrade NaN to -1e8

Config

Minimal Config Example

metric_configs:
  depth_ssim:
    pipe_name: depth-anything-v2-pipe
    default_config: ${pipes_default.depth-anything-v2-pipe}
    init_config:
    scope: edit_area
    runtime_params:
      win_size: 11
      win_sigma: 1.5
      resize_depth_maps: true

Failure Mode

Failure Semantics

The module raises a ValueError for unsupported metrics.

Other important preconditions are implicit:

• model_path must be valid at initialization
• the predicted depth map should have non-degenerate range for stable normalization

The current _normalize_depth_map(...) implementation does not explicitly guard against a zero depth range, so extremely degenerate inputs rely on inherited SSIM fallback behavior.

Extension

Extension Notes

• Extend this pipe when the comparison remains "predict depth, then compare structure."
• If you want a different depth model but the same scoring logic, the change belongs mostly in DepthAnythingv2Mixin.
• If you want a different depth-space metric altogether, add a new metric branch and document the score direction clearly.