Add README previews from sample dataset and preview generator script

Replace placeholder asset paths with five PNGs built from sample/images and
sample/labels: each composite stitches three frames edge-to-edge with YOLO
boxes and class names. Add scripts/make_readme_preview.py and preview_manifest
for reproducibility. Ignore .venv for local Pillow/PyYAML installs.

Made-with: Cursor

.gitignore

@@ -1,3 +1,6 @@
+# Local virtualenv for scripts (e.g. `scripts/make_readme_preview.py`).
+.venv/
+
 # Logs and local upload state (optional to ignore in git).
 zip_bigsplit.log
 hf_push.log

README.md

@@ -61,26 +61,25 @@ When the full dataset is present, `bigsplit/data.yaml` uses the same `path` / `t
 
 ## Preview (bounding boxes)
 
-Five sample images are chosen for dense labels and class variety, drawn with **boxes and class names on the boxes only** (no panel captions). They are stitched **edge-to-edge** at equal height — **no white padding** between frames.
-
-![Dataset preview — five annotated frames stitched](assets/README_preview.png)
-
-### Individual panels (same annotations, stable paths)
+Five PNGs are generated from the checked-in **`sample/`** images and YOLO labels. Each file stitches **three** frames **edge-to-edge** at equal height (**no padding** between panels). Boxes and class names are drawn from the corresponding `sample/labels/*.txt` files.
 
 <p align="center">
-  <img src="assets/preview_panels/panel_01.png" width="19%" alt="Panel 1" />
-  <img src="assets/preview_panels/panel_02.png" width="19%" alt="Panel 2" />
-  <img src="assets/preview_panels/panel_03.png" width="19%" alt="Panel 3" />
-  <img src="assets/preview_panels/panel_04.png" width="19%" alt="Panel 4" />
-  <img src="assets/preview_panels/panel_05.png" width="19%" alt="Panel 5" />
+  <img src="assets/preview_01.png" width="32%" alt="Sample preview 1 — three stitched frames" />
+  <img src="assets/preview_02.png" width="32%" alt="Sample preview 2 — three stitched frames" />
+  <img src="assets/preview_03.png" width="32%" alt="Sample preview 3 — three stitched frames" />
+</p>
+<p align="center">
+  <img src="assets/preview_04.png" width="32%" alt="Sample preview 4 — three stitched frames" />
+  <img src="assets/preview_05.png" width="32%" alt="Sample preview 5 — three stitched frames" />
 </p>
 
-[panel_01.png](assets/preview_panels/panel_01.png) · [panel_02.png](assets/preview_panels/panel_02.png) · [panel_03.png](assets/preview_panels/panel_03.png) · [panel_04.png](assets/preview_panels/panel_04.png) · [panel_05.png](assets/preview_panels/panel_05.png) · [manifest.txt](assets/preview_panels/manifest.txt) (source stems per panel)
+[preview_01.png](assets/preview_01.png) · [preview_02.png](assets/preview_02.png) · [preview_03.png](assets/preview_03.png) · [preview_04.png](assets/preview_04.png) · [preview_05.png](assets/preview_05.png) · [preview_manifest.txt](assets/preview_manifest.txt) (source filenames per composite)
 
 Regenerate after changing `sample/`:
 
 ```bash
-uv pip install pillow matplotlib pyyaml --python .venv/bin/python   # once
+python3 -m venv .venv   # once
+.venv/bin/pip install pillow pyyaml
 .venv/bin/python scripts/make_readme_preview.py
 ```
 

assets/preview_manifest.txt

@@ -0,0 +1,5 @@
+preview_01.png: train_images_design_1380_2_jpg.rf.a085cfddcac9b3bf7f05885498364bff.jpg + 000000546983_copy2.jpg + wider_2136.jpg
+preview_02.png: 000000053900.jpg + 000000356330_copy1.jpg + wider_2880.jpg
+preview_03.png: damagedfoodpackagingbox51_jpeg.rf.ca5fa9969495ada8fbb1922aba118445_fb8d6882.jpg + 00e9054afa5b4121.jpg + 4dfac268-175Packageeee_jpg.rf.1e33a72a2942f428c84e6c7ae40557d0_b498a222.jpg
+preview_04.png: 000000160974.jpg + N0915A42ABEX_jpg.rf.0cde92af55966ae6ac742b28930c98b3_bf12b7a7.jpg + 000000379246.jpg
+preview_05.png: 000000439141_copy1.jpg + 000032_jpg.rf.75f30b98ee2798bf1357eadb7e563345_af985b55.jpg + 000000302026.jpg

scripts/make_readme_preview.py

@@ -0,0 +1,222 @@
+#!/usr/bin/env python3
+"""
+Build README preview assets: five PNGs, each stitching three `sample/` images
+edge-to-edge at equal height with YOLO boxes and class names.
+
+Run from repo root:
+  .venv/bin/python scripts/make_readme_preview.py
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import yaml
+from PIL import Image, ImageDraw, ImageFont
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SAMPLE = REPO_ROOT / "sample"
+IMAGES_DIR = SAMPLE / "images"
+LABELS_DIR = SAMPLE / "labels"
+DATA_YAML = SAMPLE / "data.yaml"
+ASSETS_DIR = REPO_ROOT / "assets"
+
+PANEL_HEIGHT = 480
+NUM_STRIPS = 5
+PANELS_PER_STRIP = 3
+
+# Distinct colors per class index (BGR-ish order tuned for visibility on photos)
+CLASS_COLORS = [
+    (255, 99, 71),
+    (50, 205, 50),
+    (30, 144, 255),
+    (255, 215, 0),
+    (238, 130, 238),
+    (255, 140, 0),
+    (0, 206, 209),
+    (255, 20, 147),
+    (173, 255, 47),
+]
+
+
+def load_class_names() -> list[str]:
+    data = yaml.safe_load(DATA_YAML.read_text())
+    return list(data["names"])
+
+
+def parse_yolo_labels(label_path: Path) -> list[tuple[int, float, float, float, float]]:
+    out: list[tuple[int, float, float, float, float]] = []
+    if not label_path.is_file():
+        return out
+    for line in label_path.read_text().splitlines():
+        line = line.strip()
+        if not line:
+            continue
+        parts = line.split()
+        if len(parts) < 5:
+            continue
+        c = int(parts[0])
+        xc, yc, w, h = map(float, parts[1:5])
+        out.append((c, xc, yc, w, h))
+    return out
+
+
+def score_stem(stem: str) -> tuple[float, int, int]:
+    """Higher is better: prefer class variety and moderate box counts."""
+    labels = parse_yolo_labels(LABELS_DIR / f"{stem}.txt")
+    if not labels:
+        return (-1.0, 0, 0)
+    classes = {b[0] for b in labels}
+    n = len(labels)
+    uc = len(classes)
+    penalty = max(0, n - 25) * 3
+    score = uc * 50 + min(n, 20) - penalty
+    return (score, uc, n)
+
+
+def find_image_path(stem: str) -> Path | None:
+    for ext in (".jpg", ".jpeg", ".png", ".webp"):
+        p = IMAGES_DIR / f"{stem}{ext}"
+        if p.is_file():
+            return p
+    return None
+
+
+def font_or_default(size: int) -> ImageFont.FreeTypeFont | ImageFont.ImageFont:
+    for path in (
+        "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",
+        "/usr/share/fonts/truetype/liberation/LiberationSans-Regular.ttf",
+    ):
+        fp = Path(path)
+        if fp.is_file():
+            return ImageFont.truetype(str(fp), size=size)
+    return ImageFont.load_default()
+
+
+def draw_panel(img_path: Path, label_path: Path, names: list[str]) -> Image.Image:
+    img = Image.open(img_path).convert("RGB")
+    w0, h0 = img.size
+    labels = parse_yolo_labels(label_path)
+    draw = ImageDraw.Draw(img)
+    font = font_or_default(13)
+
+    for cls_id, xc, yc, bw, bh in labels:
+        if cls_id < 0 or cls_id >= len(names):
+            continue
+        x1 = (xc - bw / 2) * w0
+        y1 = (yc - bh / 2) * h0
+        x2 = (xc + bw / 2) * w0
+        y2 = (yc + bh / 2) * h0
+        color = CLASS_COLORS[cls_id % len(CLASS_COLORS)]
+        draw.rectangle([x1, y1, x2, y2], outline=color, width=2)
+        name = names[cls_id]
+        ty = max(0, y1 - 14)
+        draw.text(
+            (x1 + 1, ty),
+            name,
+            fill=color,
+            font=font,
+            stroke_width=1,
+            stroke_fill=(0, 0, 0),
+        )
+
+    # Resize to panel height
+    scale = PANEL_HEIGHT / h0
+    new_w = max(1, int(round(w0 * scale)))
+    return img.resize((new_w, PANEL_HEIGHT), Image.Resampling.LANCZOS)
+
+
+def stitch_horizontal(panels: list[Image.Image]) -> Image.Image:
+    total_w = sum(p.size[0] for p in panels)
+    h = panels[0].size[1]
+    out = Image.new("RGB", (total_w, h))
+    x = 0
+    for p in panels:
+        out.paste(p, (x, 0))
+        x += p.size[0]
+    return out
+
+
+def select_stems() -> list[str]:
+    stems: list[str] = []
+    for p in LABELS_DIR.glob("*.txt"):
+        stems.append(p.stem)
+
+    scored: list[tuple[float, int, int, str]] = []
+    for stem in stems:
+        sc, uc, n = score_stem(stem)
+        if sc < 0:
+            continue
+        if find_image_path(stem) is None:
+            continue
+        scored.append((sc, uc, n, stem))
+
+    scored.sort(key=lambda t: (t[0], t[1], t[2]), reverse=True)
+
+    # Spread selection across the ranked list so strips are not identical.
+    need = NUM_STRIPS * PANELS_PER_STRIP
+    picked: list[str] = []
+    n_ranked = len(scored)
+    if n_ranked == 0:
+        raise SystemExit("No labeled images found under sample/.")
+
+    step = max(1, n_ranked // need)
+    i = 0
+    for _ in range(need):
+        _, _, _, stem = scored[i % n_ranked]
+        picked.append(stem)
+        i += step
+
+    # Deduplicate while preserving order; back-fill from top scores
+    seen: set[str] = set()
+    unique: list[str] = []
+    for s in picked:
+        if s not in seen:
+            seen.add(s)
+            unique.append(s)
+
+    for _, _, _, stem in scored:
+        if len(unique) >= need:
+            break
+        if stem not in seen:
+            seen.add(stem)
+            unique.append(stem)
+
+    if len(unique) < need:
+        raise SystemExit(f"Need {need} distinct images; only found {len(unique)}.")
+
+    return unique[:need]
+
+
+def main() -> None:
+    names = load_class_names()
+    stems = select_stems()
+    ASSETS_DIR.mkdir(parents=True, exist_ok=True)
+
+    manifest_lines: list[str] = []
+
+    for strip_idx in range(NUM_STRIPS):
+        chunk = stems[strip_idx * PANELS_PER_STRIP : (strip_idx + 1) * PANELS_PER_STRIP]
+        panels: list[Image.Image] = []
+        for stem in chunk:
+            img_p = find_image_path(stem)
+            assert img_p is not None
+            lbl_p = LABELS_DIR / f"{stem}.txt"
+            panels.append(draw_panel(img_p, lbl_p, names))
+
+        strip = stitch_horizontal(panels)
+        out_path = ASSETS_DIR / f"preview_{strip_idx + 1:02d}.png"
+        strip.save(out_path, optimize=True)
+        parts = []
+        for s in chunk:
+            ip = find_image_path(s)
+            assert ip is not None
+            parts.append(ip.name)
+        manifest_lines.append(f"{out_path.name}: {' + '.join(parts)}")
+
+    (ASSETS_DIR / "preview_manifest.txt").write_text("\n".join(manifest_lines) + "\n")
+    print(f"Wrote {NUM_STRIPS} previews to {ASSETS_DIR}/")
+
+
+if __name__ == "__main__":
+    main()