scripts/compute_metrics.py

#!/usr/bin/env python
"""
Batch wrapper around EvalMDE's metrics.

Dual-track output per scene:
  RAW   — pred fed verbatim into SAWA-H components (EvalMDE official protocol;
          biased against affine-invariant models in boundary_F1 / rel_normal
          since those metrics are NOT shift-invariant)
  ALIGNED — pred affine-aligned to GT (LSQ fit on valid mask) before feeding,
          gives fair cross-model comparison

Both tracks emit the 5 SAWA-H components + the weighted sum + raw rel_normal.
"""
import json, click
from pathlib import Path
import numpy as np
import torch

from evalmde.utils.depth import load_data, align
from evalmde.utils.proj import depth_to_xyz
from evalmde.utils.torch import reformat_as_torch_tensor
from evalmde.metrics.standard import rel_depth, delta0125
from evalmde.metrics.boundary import boundary_f1
from evalmde.metrics.rel_normal import rel_normal as rel_normal_fn

# SAWA-H weights (verbatim from evalmde/metrics/sawa_h.py)
W_WKDR, W_DDISP_AF, W_DDEPTH_AF, W_BOUND_F1, W_RELNORM = 3.65, 0.18, 0.01, 0.20, 1.94


def compute_components(pr_depth, pr_intr, pr_valid, gt_depth, gt_intr, gt_valid):
    """Verbatim of sawa_h.compute_sawa_h body, returns 5 comps + weighted sum."""
    wkdr = 1.0 - rel_depth(pr_depth, gt_depth, gt_valid)[1]

    d_disp = 1.0 - delta0125(
        align(1 / reformat_as_torch_tensor(pr_depth),
              reformat_as_torch_tensor(gt_depth),
              reformat_as_torch_tensor(gt_valid),
              'disparity_affine_clip_by_0'),
        gt_depth, gt_valid)[1]

    d_depth = 1.0 - delta0125(
        align(reformat_as_torch_tensor(pr_depth),
              reformat_as_torch_tensor(gt_depth),
              reformat_as_torch_tensor(gt_valid),
              'depth_affine_lst_sq_clip_by_0'),
        gt_depth, gt_valid)[1]

    bound = boundary_f1(reformat_as_torch_tensor(pr_depth),
                        reformat_as_torch_tensor(gt_depth),
                        reformat_as_torch_tensor(gt_valid))[1]

    relnorm = rel_normal_fn(
        depth_to_xyz(gt_intr, gt_depth), gt_valid,
        depth_to_xyz(pr_intr, pr_depth), pr_valid,
    )

    sawa = (W_WKDR * wkdr + W_DDISP_AF * d_disp + W_DDEPTH_AF * d_depth
            + W_BOUND_F1 * bound + W_RELNORM * relnorm)

    return {
        'wkdr_no_align':                  float(wkdr),
        'delta0125_disparity_affine_err': float(d_disp),
        'delta0125_depth_affine_err':     float(d_depth),
        'boundary_f1_err':                float(bound),
        'rel_normal':                     float(relnorm),
        'sawa_h':                         float(sawa),
    }


def affine_align_to_gt(pr_d, gt_d, valid):
    """LSQ-fit (a, b) so that a*pr + b ≈ gt on valid pixels."""
    m = valid & np.isfinite(pr_d) & np.isfinite(gt_d) & (np.abs(pr_d) < 1e6)
    if m.sum() < 100:
        return pr_d.astype(np.float32), 1.0, 0.0
    p = pr_d[m].astype(np.float64)
    g = gt_d[m].astype(np.float64)
    A = np.stack([p, np.ones_like(p)], axis=-1)
    a, b = np.linalg.lstsq(A, g, rcond=None)[0]
    aligned = (pr_d.astype(np.float64) * a + b).clip(1e-4, None)
    return aligned.astype(np.float32), float(a), float(b)


@click.command()
@click.option('--gt-root',    required=True, type=click.Path())
@click.option('--pred-root',  required=True, type=click.Path())
@click.option('--model-name', required=True, type=str)
@click.option('--output',     required=True, type=click.Path())
def main(gt_root, pred_root, model_name, output):
    gt_root, pred_root = Path(gt_root), Path(pred_root) / model_name
    scenes = sorted(d.name for d in pred_root.iterdir() if (d / 'pred_depth.npz').exists())
    print(f'Found {len(scenes)} scenes for {model_name}')

    rows = []
    for sc in scenes:
        gt_d, gt_i, gt_v = load_data(str(gt_root  / sc / 'gt_depth.npz'))
        pr_d, pr_i, pr_v = load_data(str(pred_root / sc / 'pred_depth.npz'))

        raw = compute_components(pr_d, pr_i, pr_v, gt_d, gt_i, gt_v)

        pr_d_aln, scale, shift = affine_align_to_gt(pr_d, gt_d, pr_v & gt_v)
        # Aligned pred uses gt intrinsics (we don't realign focal — only depth values)
        aln = compute_components(pr_d_aln, gt_i, pr_v, gt_d, gt_i, gt_v)
        aln['_align_scale'] = scale
        aln['_align_shift'] = shift

        rows.append({'scene': sc, 'raw': raw, 'aligned': aln})
        print(f'  {sc}: '
              f'sawa_h raw={raw["sawa_h"]:.3f} aln={aln["sawa_h"]:.3f}  |  '
              f'relnorm raw={raw["rel_normal"]:.3f} aln={aln["rel_normal"]:.3f}  |  '
              f'boundF1_err raw={raw["boundary_f1_err"]:.3f} aln={aln["boundary_f1_err"]:.3f}')

    n = max(len(rows), 1)
    keys = ['wkdr_no_align','delta0125_disparity_affine_err','delta0125_depth_affine_err',
            'boundary_f1_err','rel_normal','sawa_h']
    mean_raw = {k: sum(r['raw'][k] for r in rows) / n for k in keys}
    mean_aln = {k: sum(r['aligned'][k] for r in rows) / n for k in keys}

    summary = {
        'model': model_name,
        'n_scenes': len(rows),
        'sawa_h_weights': {
            'wkdr_no_align': W_WKDR, 'delta0125_disparity_affine_err': W_DDISP_AF,
            'delta0125_depth_affine_err': W_DDEPTH_AF,
            'boundary_f1_err': W_BOUND_F1, 'rel_normal': W_RELNORM,
        },
        'per_scene': rows,
        'mean_raw':     mean_raw,
        'mean_aligned': mean_aln,
    }
    Path(output).parent.mkdir(parents=True, exist_ok=True)
    json.dump(summary, open(output, 'w'), indent=2)
    print(f'\nMean RAW    : {mean_raw}')
    print(f'Mean ALIGNED: {mean_aln}')
    print(f'Saved → {output}')


if __name__ == '__main__':
    main()