hasari-api / scripts /prepare_cardd_from_json.py
erdoganpeker's picture
v0.3.0 — multimodal vehicle damage MVP
e327f0d
"""
prepare_cardd_from_json.py — CarDD HF mirror samples.json → YOLO segmentation format
Fast direct parser. No FiftyOne dependency. Reads samples.json (FiftyOne export format)
and converts each detection's mask (zlib-compressed binary, base64) to YOLO segmentation
polygons.
Usage:
python scripts/prepare_cardd_from_json.py \\
--samples C:\\Users\\Erdogan\\fiftyone\\huggingface\\hub\\harpreetsahota\\CarDD\\samples.json \\
--images_root C:\\Users\\Erdogan\\Desktop\\arac-hasar-v2\\services\\ml\\data\\cardd_hf \\
--output_dir services/ml/data/cardd_yolo
"""
from __future__ import annotations
import argparse
import base64
import io
import json
import random
import shutil
import sys
import zlib
from pathlib import Path
import numpy as np
import cv2
from tqdm import tqdm
# Class mapping: label (lowercase) → class_id (matches services/ml/cardd.yaml)
CLASS_MAP = {
"dent": 0,
"scratch": 1,
"crack": 2,
"glass shatter": 3,
"lamp broken": 4,
"tire flat": 5,
}
CLASS_NAMES = ["dent", "scratch", "crack", "glass_shatter", "lamp_broken", "tire_flat"]
def decode_mask(mask_b64: str) -> np.ndarray | None:
"""FiftyOne mask: zlib-compressed numpy .npy bytes, base64-encoded.
Returns a 2D boolean (or uint8) array, or None if decode fails.
"""
try:
raw = base64.b64decode(mask_b64)
decompressed = zlib.decompress(raw)
arr = np.load(io.BytesIO(decompressed))
return arr
except Exception:
return None
def mask_to_polygon(
mask_arr: np.ndarray,
bbox: list, # [x_min, y_min, w, h] normalized
img_w: int,
img_h: int,
) -> list[list[tuple[float, float]]]:
"""Convert mask + bbox to normalized image-level polygons.
Mask is a 2D array (H, W) at bbox pixel resolution.
"""
if mask_arr is None or mask_arr.ndim != 2:
return []
bx, by, _, _ = bbox
mh, mw = mask_arr.shape
mask_2d = mask_arr
# Find contours in mask
contours, _ = cv2.findContours(
(mask_2d > 0).astype(np.uint8) * 255,
cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE,
)
polygons = []
for c in contours:
if len(c) < 3:
continue
# Simplify polygon to reduce points (epsilon = 0.5% of perimeter)
peri = cv2.arcLength(c, True)
approx = cv2.approxPolyDP(c, 0.005 * peri, True)
if len(approx) < 3:
continue
# Convert from mask-local pixels to image-normalized coords
poly = []
for pt in approx.squeeze(axis=1):
px, py = pt[0], pt[1]
# mask-local pixel → image-pixel: image_x = bx*img_w + px
image_x = bx * img_w + px
image_y = by * img_h + py
nx = image_x / img_w
ny = image_y / img_h
poly.append((max(0.0, min(1.0, nx)), max(0.0, min(1.0, ny))))
polygons.append(poly)
return polygons
def bbox_to_polygon(bbox: list) -> list[tuple[float, float]]:
"""Fallback: bbox as 4-point polygon (normalized)."""
bx, by, bw, bh = bbox
return [(bx, by), (bx + bw, by), (bx + bw, by + bh), (bx, by + bh)]
def process_sample(sample: dict, images_root: Path, splits: dict, out_dir: Path) -> str:
"""Returns: 'ok' | 'skip_no_img' | 'skip_no_label' | 'skip_decode_fail'"""
rel_path = sample["filepath"] # e.g., "data/000001.jpg"
img_path = images_root / rel_path
if not img_path.exists():
return "skip_no_img"
meta = sample.get("metadata") or {}
img_w = meta.get("width", 0)
img_h = meta.get("height", 0)
if not (img_w and img_h):
# Read from file
from PIL import Image
with Image.open(img_path) as im:
img_w, img_h = im.size
segmentations = sample.get("segmentations") or {}
detections_list = segmentations.get("detections") or []
# Fallback to detections if no segmentations
if not detections_list:
det_field = sample.get("detections") or {}
detections_list = det_field.get("detections") or []
yolo_lines = []
for det in detections_list:
label = (det.get("label") or "").lower().strip()
if label not in CLASS_MAP:
continue
cls_id = CLASS_MAP[label]
bbox = det.get("bounding_box") or []
if len(bbox) != 4:
continue
polygons = []
mask_obj = det.get("mask")
if mask_obj and "$binary" in mask_obj:
b64 = mask_obj["$binary"].get("base64", "")
if b64:
try:
mask_arr = decode_mask(b64)
polygons = mask_to_polygon(mask_arr, bbox, img_w, img_h)
except Exception:
polygons = []
if not polygons:
# Fallback to bbox
polygons = [bbox_to_polygon(bbox)]
for poly in polygons:
if len(poly) < 3:
continue
coords = " ".join(f"{x:.6f} {y:.6f}" for (x, y) in poly)
yolo_lines.append(f"{cls_id} {coords}")
if not yolo_lines:
return "skip_no_label"
sid = sample.get("_id", {}).get("$oid", "")
split = splits.get(sid, "train")
# Copy/hardlink image
dst_img = out_dir / "images" / split / img_path.name
if not dst_img.exists():
try:
import os
os.link(img_path, dst_img) # Hardlink for zero-copy
except OSError:
shutil.copy2(img_path, dst_img)
# Write labels
label_path = out_dir / "labels" / split / (img_path.stem + ".txt")
label_path.write_text("\n".join(yolo_lines), encoding="utf-8")
return "ok"
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--samples", type=Path, required=True)
ap.add_argument("--images_root", type=Path, required=True,
help="Path containing 'data/' subdir with images")
ap.add_argument("--output_dir", type=Path, required=True)
ap.add_argument("--train_ratio", type=float, default=0.8)
ap.add_argument("--val_ratio", type=float, default=0.1)
ap.add_argument("--seed", type=int, default=42)
args = ap.parse_args()
print(f">> Loading samples.json: {args.samples}")
with open(args.samples, "r", encoding="utf-8") as f:
data = json.load(f)
samples = data.get("samples", data) if isinstance(data, dict) else data
print(f">> Loaded {len(samples)} samples")
# Prepare output dirs
out = args.output_dir
for sp in ("train", "val", "test"):
(out / "images" / sp).mkdir(parents=True, exist_ok=True)
(out / "labels" / sp).mkdir(parents=True, exist_ok=True)
# Split
random.seed(args.seed)
sample_ids = [s.get("_id", {}).get("$oid", f"id_{i}") for i, s in enumerate(samples)]
shuffled = list(zip(sample_ids, samples))
random.shuffle(shuffled)
n = len(shuffled)
n_train = int(n * args.train_ratio)
n_val = int(n * args.val_ratio)
splits = {}
for i, (sid, _) in enumerate(shuffled):
if i < n_train:
splits[sid] = "train"
elif i < n_train + n_val:
splits[sid] = "val"
else:
splits[sid] = "test"
counts = {"ok": 0, "skip_no_img": 0, "skip_no_label": 0, "skip_decode_fail": 0}
for sample in tqdm(samples, desc="Converting"):
result = process_sample(sample, args.images_root, splits, out)
counts[result] = counts.get(result, 0) + 1
print()
print(">> Conversion complete:")
for k, v in counts.items():
print(f" {k}: {v}")
# Per-split tallies
print(">> Per-split tallies:")
for sp in ("train", "val", "test"):
n_imgs = len(list((out / "images" / sp).iterdir()))
n_lbls = len(list((out / "labels" / sp).iterdir()))
print(f" {sp}: images={n_imgs}, labels={n_lbls}")
# Write YAML config
yaml_path = out / "cardd.yaml"
yaml_path.write_text(
"# Auto-generated by prepare_cardd_from_json.py\n"
f"path: {out.resolve().as_posix()}\n"
"train: images/train\n"
"val: images/val\n"
"test: images/test\n"
"nc: 6\n"
"names:\n" + "\n".join(f" {i}: {n}" for i, n in enumerate(CLASS_NAMES)) + "\n",
encoding="utf-8",
)
print(f">> Wrote dataset config: {yaml_path}")
if __name__ == "__main__":
main()