visualize_labels.py

#!/usr/bin/env python3
"""Visualize SPARK 2022 dataset labels.

Dataset layout:
    labels/{train,val,test}.csv   with rows:  filename,class,bbox
    {train,val,test}/             image folders (.jpg)

bbox = [xmin, ymin, xmax, ymax]   (x = column, y = row; origin = top-left)

Plots images with their bounding box and class label, in grids of up to
6 per figure. Figures are saved as PNG next to this script (headless-safe)
and also shown in a window when a display is available.

Examples:
    python3 visualize_labels.py                          # 6 random from train
    python3 visualize_labels.py val --num 12 --seed 3    # reproducible sample
    python3 visualize_labels.py test --filenames img057676.jpg,img058116.jpg
    python3 visualize_labels.py train --class smart_1    # sample one class
    python3 visualize_labels.py val --save /tmp/out.png
"""

import argparse
import ast
import csv
import os
import random
import sys
from pathlib import Path

import matplotlib
if not os.environ.get("DISPLAY") and sys.platform != "darwin":
    matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle
from PIL import Image

BOX_COLOR = "#00FF66"       # bright green — visible on dark space imagery
PER_FIGURE = 6              # images per figure, 3 columns


def load_labels(csv_path: Path):
    """Read the CSV into {filename: (class, [xmin, ymin, xmax, ymax])}."""
    rows = {}
    with open(csv_path, newline="") as f:
        reader = csv.reader(f)
        next(reader)  # header: filename,class,bbox
        for row in reader:
            if len(row) != 3:
                continue
            filename, cls, bbox_raw = row
            try:
                bbox = list(ast.literal_eval(bbox_raw))
            except (ValueError, SyntaxError):
                bbox = None
            rows[filename] = (cls, bbox)
    return rows


def plot_batch(batch, rows, img_dir: Path, out_path: Path):
    """Draw up to PER_FIGURE labeled images on one figure and save it."""
    ncols = min(3, len(batch))
    nrows = -(-len(batch) // ncols)
    fig, axes = plt.subplots(nrows, ncols, figsize=(5 * ncols, 5 * nrows),
                             squeeze=False)
    for ax in axes.flat:
        ax.set_axis_off()

    for ax, name in zip(axes.flat, batch):
        cls, bbox = rows[name]
        img_path = img_dir / name
        if not img_path.exists():
            ax.set_title(f"{name}\nIMAGE NOT FOUND", fontsize=9, color="red")
            continue
        with Image.open(img_path) as im:
            ax.imshow(im)
        if bbox is None or len(bbox) != 4:
            ax.set_title(f"{name} — {cls}\nBAD BBOX", fontsize=9, color="red")
            continue
        xmin, ymin, xmax, ymax = bbox
        ax.add_patch(Rectangle((xmin, ymin), xmax - xmin, ymax - ymin,
                               linewidth=2, edgecolor=BOX_COLOR,
                               facecolor="none"))
        ax.text(xmin, max(ymin - 6, 0), cls, fontsize=9, color="black",
                va="bottom", ha="left",
                bbox=dict(facecolor=BOX_COLOR, edgecolor="none",
                          boxstyle="round,pad=0.2"))
        ax.set_title(f"{name} — {cls}", fontsize=9)

    fig.tight_layout()
    fig.savefig(out_path, dpi=100, bbox_inches="tight")
    print(f"saved: {out_path}")
    return fig


def main():
    ap = argparse.ArgumentParser(
        description="Plot dataset images with their [xmin, ymin, xmax, ymax] "
                    "bounding boxes.")
    ap.add_argument("split", nargs="?", default="train",
                    choices=["train", "val", "test"],
                    help="which split to visualize (default: train)")
    ap.add_argument("--root", type=Path, default=Path(__file__).resolve().parent,
                    help="dataset root (default: this script's folder)")
    ap.add_argument("--num", type=int, default=6,
                    help="number of random images to plot (default: 6)")
    ap.add_argument("--filenames",
                    help="comma-separated filenames to plot instead of a "
                         "random sample")
    ap.add_argument("--class", dest="cls",
                    help="restrict the random sample to one class")
    ap.add_argument("--seed", type=int, default=None,
                    help="random seed, for a reproducible sample")
    ap.add_argument("--save", type=Path, default=None,
                    help="output PNG path (default: preview_<split>.png in --root)")
    args = ap.parse_args()

    csv_path = args.root / "labels" / f"{args.split}.csv"
    img_dir = args.root / args.split
    if not csv_path.is_file():
        sys.exit(f"labels file not found: {csv_path}")

    rows = load_labels(csv_path)

    if args.filenames:
        stem_to_name = {Path(n).stem: n for n in rows}
        picked, missing = [], []
        for n in (s.strip() for s in args.filenames.split(",")):
            resolved = n if n in rows else stem_to_name.get(Path(n).stem)
            picked.append(resolved) if resolved else missing.append(n)
        if missing:
            sys.exit(f"not found in {csv_path.name}: {missing}")
    else:
        pool = sorted(n for n, (cls, _) in rows.items()
                      if args.cls is None or cls == args.cls)
        if not pool:
            sys.exit(f"no rows for class {args.cls!r} in {csv_path.name}")
        picked = random.Random(args.seed).sample(pool, min(args.num, len(pool)))

    base = args.save or (args.root / f"preview_{args.split}.png")
    figs = []
    for i in range(0, len(picked), PER_FIGURE):
        suffix = f"_{i // PER_FIGURE + 1}" if len(picked) > PER_FIGURE else ""
        out = base.with_name(f"{base.stem}{suffix}.png")
        figs.append(plot_batch(picked[i:i + PER_FIGURE], rows, img_dir, out))

    if matplotlib.get_backend().lower() != "agg":
        plt.show()
    else:
        plt.close("all")


if __name__ == "__main__":
    main()