jamiewjm/sam-tp

SAM‑TP Traversability Dataset

This repository contains pixel‑wise traversability masks paired with egocentric RGB images, prepared in a flat, filename‑aligned layout that is convenient for training SAM‑2 / SAM‑TP‑style segmentation models.

Folder layout

.
├─ images/          # RGB frames (.jpg/.png). Filenames are globally unique.
├─ annotations/     # Binary masks (.png/.jpg). Filenames match images 1‑to‑1.
└─ manifest.csv     # Provenance rows and any missing‑pair notes.

Each annotations/<FILENAME> is the mask for images/<FILENAME> (same filename, different folder).

---

File naming

Filenames are made globally unique by concatenating the original subfolder path and the local stem with __ separators, e.g.

ride_68496_8ef98b_20240716023032_517__1.jpg
ride_68496_8ef98b_20240716023032_517__1.png  # corresponding mask

---

Mask format

• Single‑channel binary masks; foreground = traversable, background = non‑traversable.
• Stored as .png or .jpg depending on source. If your pipeline requires PNG, convert on the fly in your dataloader.
• Values are typically {0, 255}. You can binarize via mask = (mask > 127).astype(np.uint8).

---

How to use

A) Load with datasets (ImageFolder‑style)

from datasets import load_dataset
from pathlib import Path
from PIL import Image

REPO = "jamiewjm/sam-tp"  # e.g. "jamiewjm/sam-tp"

ds_imgs = load_dataset(
    "imagefolder",
    data_dir=".",
    data_files={"image": f"hf://datasets/{REPO}/images/**"},
    split="train",
)
ds_msks = load_dataset(
    "imagefolder",
    data_dir=".",
    data_files={"mask": f"hf://datasets/{REPO}/annotations/**"},
    split="train",
)

# Build a mask index by filename
mask_index = {Path(r["image"]["path"]).name: r["image"]["path"] for r in ds_msks}

row = ds_imgs[0]
img_path = Path(row["image"]["path"])
msk_path = Path(mask_index[img_path.name])

img = Image.open(img_path).convert("RGB")
msk = Image.open(msk_path).convert("L")

B) Minimal PyTorch dataset

from pathlib import Path
from PIL import Image
from torch.utils.data import Dataset

class TraversabilityDataset(Dataset):
    def __init__(self, root):
        root = Path(root)
        self.img_dir = root / "images"
        self.msk_dir = root / "annotations"
        self.items = sorted([p for p in self.img_dir.iterdir() if p.is_file()])
    def __len__(self):
        return len(self.items)
    def __getitem__(self, idx):
        ip = self.items[idx]
        mp = self.msk_dir / ip.name
        return Image.open(ip).convert("RGB"), Image.open(mp).convert("L")

C) Pre‑processing notes for SAM‑2/SAM‑TP training

• Resize/pad to your training resolution (commonly 1024×1024) with masks aligned.
• Normalize images per your backbone’s recipe.
• If your trainer expects COCO‑RLE masks, convert PNG → RLE in the dataloader stage.

---

Provenance & splits

• The dataset was flattened from mirrored directory trees (images and annotations) with 1‑to‑1 filename alignment.
• If you create explicit train/val/test splits, please add a split column to a copy of manifest.csv and contribute it back.

---

License

Data: CC‑BY‑4.0 (Attribution). See LICENSE for details.

---

Citation

If you use this dataset in academic or industrial research, please cite the accompanying paper/report describing the data collection and labeling protocol:

GeNIE: A Generalizable Navigation System for In-the-Wild Environments

Available at: https://arxiv.org/abs/2506.17960

Contains the SAM-TP traversability dataset and evaluation methodology.

@article{wang2025genie,
  title   = {GeNIE: A Generalizable Navigation System for In-the-Wild Environments},
  author  = {Wang, Jiaming and et al.},
  journal = {arXiv preprint arXiv:2506.17960},
  year    = {2025},
  url     = {https://arxiv.org/abs/2506.17960}
}