NeurIPS_2026_BLV / tools /evaluate.py
NavAble's picture
Dataset card: paper, authors, tools and figures
ece131e verified
Raw History Blame Contribute Delete
3.82 kB
#!/usr/bin/env python3
"""Score segmentation predictions on a real_ours split with the paper's metrics.
Predictions are label-map PNGs in --pred-dir, one per image, named like the image
(`row["image"]["path"]`, e.g. 0f4eab9c-frame_003200.png) and the same size as the image.
Reported (in %): mIoU, mPrecision and mRecall over background plus the 10 classes with real-world
examples, as in the paper's tables, the same means over the 10 classes only (fg_*), and per-class
values. revolving_door has no real-world examples and is not scored.
Examples:
python tools/evaluate.py --pred-dir preds/ # dataset class ids
python tools/evaluate.py --pred-dir preds/ --label-order checkpoint # raw output of a released mmseg checkpoint
python tools/evaluate.py --pred-dir preds/ --gt-dir data/navable/ann_dir/test
"""
from __future__ import annotations
import argparse
import json
import sys
from pathlib import Path
import numpy as np
from PIL import Image
sys.path.insert(0, str(Path(__file__).resolve().parent))
import navable as nv # noqa: E402
def ground_truth(args):
"""Yield (file name, label map) for every image of the split."""
if args.gt_dir:
for f in sorted(args.gt_dir.glob("*.png")):
yield f.name, nv.mask_array(Image.open(f))
return
files = nv.parquet_files("real_ours", args.split, args.data_dir)
for row in nv.iter_rows(files, columns=["image.path", "mask"]):
yield Path(nv.row_name(row)).stem + ".png", nv.mask_array(row["mask"])
def print_table(r: dict) -> None:
print(f"\n{'class':<20}{'IoU':>8}{'Prec':>8}{'Rec':>8}")
for name in nv.CLASSES:
c = r["per_class"][name]
if c["gt_pixels"] or name in ("background",):
print(f"{name:<20}{c['IoU']:8.2f}{c['precision']:8.2f}{c['recall']:8.2f}")
print(f"\n{'paper (bg + 10)':<20}{r['mIoU']:8.2f}{r['mPrecision']:8.2f}{r['mRecall']:8.2f}")
print(f"{'10 classes only':<20}{r['fg_mIoU']:8.2f}{r['fg_mPrecision']:8.2f}{r['fg_mRecall']:8.2f}")
def main() -> int:
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--pred-dir", type=Path, required=True)
p.add_argument("--split", choices=nv.CONFIGS["real_ours"], default="test")
p.add_argument("--label-order", choices=["dataset", "checkpoint", "checkpoint-fg"], default="dataset",
help="label ids in the predictions: dataset class ids, a released mmseg checkpoint's "
"order (12 labels), or SAN's order without background (11 labels)")
p.add_argument("--gt-dir", type=Path, help="ground-truth label PNGs (e.g. from tools/export.py) instead of the Hub")
p.add_argument("--data-dir", type=Path, help="downloaded copy of the repo (see tools/download.py)")
p.add_argument("--out", type=Path, help="write the full results as JSON")
args = p.parse_args()
metrics, missing, n = nv.SegmentationMetrics(), [], 0
for name, gt in ground_truth(args):
f = args.pred_dir / name
if not f.exists():
missing.append(name)
continue
pred = np.asarray(Image.open(f))
if args.label_order != "dataset":
pred = nv.checkpoint_to_dataset_labels(pred, has_background=args.label_order == "checkpoint")
metrics.update(pred, gt)
n += 1
if missing:
print(f"{len(missing)} images have no prediction in {args.pred_dir} (e.g. {missing[:3]})")
return 1
r = metrics.results()
print(f"real_ours/{args.split}: {n} images")
print_table(r)
if args.out:
args.out.write_text(json.dumps({"split": args.split, "images": n, **r}, indent=1))
print(f"\nresults -> {args.out}")
return 0
if __name__ == "__main__":
sys.exit(main())