SteelDS (Parquet mirror)

Shredded E40 steel scrap and copper on a lab conveyor belt, 3680 x 2160 frames with YOLO segmentation labels. Mirror of https://zenodo.org/records/20271102, converted for streaming (videos left out). Licence: CC BY 4.0, credit the original authors.

• Subsets (configs): a1, a2, a3; a1-a3 are labelled, a4-a5 unlabelled
• Classes: 0 = background, 1 = steel, 2 = copper
• yolo_label: the original YOLO-seg label text (" x1 y1 x2 y2 ..." normalised to 0-1); has_label = False where the original frame had no label file

Load

from datasets import load_dataset

ds = load_dataset("vivekkol/steelds", "a1", split="train", streaming=True)   # or without streaming=True to download once
for row in ds:
    image = row["image"]          # PIL image
    label = row["yolo_label"]     # YOLO-seg text, same as the original .txt file

Write out as YOLO folders (for Ultralytics)

import os
from datasets import load_dataset

for split in ["train", "val", "test"]:
    os.makedirs(f"steelds/images/{split}", exist_ok=True); os.makedirs(f"steelds/labels/{split}", exist_ok=True)
    for row in load_dataset("vivekkol/steelds", "a1", split=split, streaming=True):
        row["image"].save(f"steelds/images/{split}/{row['file_name']}.jpg", quality=95)
        open(f"steelds/labels/{split}/{row['file_name']}.txt", "w").write(row["yolo_label"])