src/model/encoder/depth/infinidepth/warp_utils.py

import torch
from einops import rearrange

EPS = 1e-6


class WarpMedian:
    def __init__(self, **kwargs):
        pass

    def warp(self, depth, **kwargs):
        if kwargs.get("reference_meta", None) is not None:
            median_val = kwargs["reference_meta"]
            return depth / torch.clamp(median_val, min=1e-3), (depth > EPS), median_val
        prompt_depth = kwargs.get("prompt_depth")
        prompt_mask = kwargs.get("prompt_mask")
        median_val = []
        batch_size = depth.shape[0]
        for b in range(batch_size):
            valid_prompt = (prompt_mask[b] > 0.0) & torch.isfinite(prompt_depth[b]) & (prompt_depth[b] > 0.0)
            if not valid_prompt.any():
                raise RuntimeError(
                    "WarpMedian received an empty prompt set. Prompt-conditioned inputs "
                    "must be filtered at the dataset stage before reaching InfiniDepth."
                )
            median = torch.quantile(prompt_depth[b][valid_prompt], 0.5)
            median_val.append(median)
        median_val = torch.stack(median_val, dim=0)
        median_val = rearrange(median_val, "b -> b 1 1 1")
        return depth / torch.clamp(median_val, min=1e-2), (depth > EPS) & (prompt_mask > 0.0), median_val

    def unwarp(self, depth, **kwargs):
        median_val = kwargs.get("reference_meta")
        return depth * torch.clamp(median_val, min=1e-3)