NavAble/NeurIPS_2026_BLV

BLV Object Recognition: Synthetic + Real-World

A dataset for training and evaluating object recognition and segmentation models on infrastructure relevant to blind and low-vision (BLV) navigation in urban environments. Three configurations plus a flat tree of 3D assets:

Config / tree	Splits	Purpose
syn	train	Photorealistic IsaacSim renders for training / pretraining.
real_ours	train / validation / test	Real photographs we captured. real_ours/test is the canonical benchmark eval.
real_curated	train	Curated frames from public HF segmentation datasets (curation, mapillary), remapped to our class palette.
synthetic_objects/ (tree)	n/a	3D asset library: per-asset .glb + .ply + .usdz triples grouped by BLV class.

Quick links

• Datasheet for Datasets
• Class index + palette
• Croissant metadata is auto-generated by Hugging Face for this repo (look for the Croissant button on the dataset page).
• Paper: NeurIPS 2026 Datasets & Benchmarks (TBD).

Loading

With datasets

from datasets import load_dataset

syn_train     = load_dataset("NavAble/NeurIPS_2026_BLV", "syn",          split="train")
ours_train    = load_dataset("NavAble/NeurIPS_2026_BLV", "real_ours",    split="train")
ours_val      = load_dataset("NavAble/NeurIPS_2026_BLV", "real_ours",    split="validation")
ours_test     = load_dataset("NavAble/NeurIPS_2026_BLV", "real_ours",    split="test")  # canonical eval
curated_train = load_dataset("NavAble/NeurIPS_2026_BLV", "real_curated", split="train")

row = ours_test[0]
row["image"]   # PIL.Image.Image, RGB
row["mask"]    # PIL.Image.Image, P-mode (palette) - pixel value == class_id

Pulling the 3D assets

from huggingface_hub import snapshot_download

# All 3D assets for a single class:
snapshot_download(
    repo_id="NavAble/NeurIPS_2026_BLV", repo_type="dataset",
    allow_patterns=["synthetic_objects/door_button/**"],
    local_dir="./assets",
)

With PyTorch directly

from torch.utils.data import Dataset
from datasets import load_dataset
import numpy as np
import torch
import torchvision.transforms.functional as TF

class BLVSegDataset(Dataset):
    def __init__(self, config: str, split: str, image_size: int = 512):
        self.ds = load_dataset("NavAble/NeurIPS_2026_BLV", config, split=split)
        self.image_size = image_size

    def __len__(self):
        return len(self.ds)

    def __getitem__(self, idx):
        row = self.ds[idx]
        img = TF.resize(row["image"].convert("RGB"), [self.image_size, self.image_size])
        mask = TF.resize(row["mask"], [self.image_size, self.image_size],
                         interpolation=TF.InterpolationMode.NEAREST)
        img_t  = TF.to_tensor(img)
        mask_t = torch.from_numpy(np.array(mask, dtype=np.int64))
        return {"image": img_t, "mask": mask_t, "class": row["object_class"]}

Splits & sizes

Config	Split	Rows
syn	train	452704
real_ours	train	3703
real_ours	validation	396
real_ours	test	1482
real_curated	train	36466

3D asset library (synthetic_objects/): 500 GLB+PLY+USDZ triples across 9 classes.

Class taxonomy

ID	Class	Synthetic	Real (Ours)
1	aps_button	yes	yes
2	bus_stop	yes	yes
3	bus_stop_sign	yes	yes
4	crosswalk	yes	yes
5	door_button	yes	yes
6	elevator	yes	yes
7	elevator_button	yes	yes
8	escalator	yes	yes
9	handrail	yes	yes
10	pedestrian_signal	yes	yes
11	turnstile	yes	no

The synthetic-only class turnstile has no real-world examples in this release; report real-world metrics over the 10 shared classes.

Per-class row counts

Class	syn/train	real_ours/train	real_ours/val	real_ours/test	real_curated/train
aps_button	62855	206	23	66	0
bus_stop	60789	205	23	62	0
bus_stop_sign	60480	140	16	54	0
crosswalk	54360	9	1	3	27786
door_button	45360	1327	148	622	0
elevator	23760	1065	119	479	15
elevator_button	23350	378	23	86	4401
escalator	7062	135	15	40	1296
handrail	44468	21	3	8	1197
pedestrian_signal	45210	217	25	62	6650
turnstile	25010	0	0	0	0

Synthetic coverage

• Frames: 452704
• Trajectories: 700
• Environments: 37
• Distinct assets: 112

Mask encoding

Each mask is a single-channel PNG (PIL mode="P") with an embedded palette. Pixel value i corresponds to the i-th entry in class_index.json:

Pixel	Class	Palette RGB
0	BACKGROUND	(0, 0, 0)
1	aps_button	(220, 20, 60)
2	bus_stop	(255, 140, 0)
3	bus_stop_sign	(255, 215, 0)
4	crosswalk	(50, 205, 50)
5	door_button	(0, 191, 255)
6	elevator	(138, 43, 226)
7	elevator_button	(255, 105, 180)
8	escalator	(0, 128, 128)
9	handrail	(165, 42, 42)
10	pedestrian_signal	(75, 0, 130)
11	turnstile	(255, 20, 147)

Convert to a numeric label map with np.array(row["mask"]).

Source data

• Synthetic — generated in NVIDIA IsaacSim (Replicator) by spawning each asset across a curated catalog of urban environments (37 unique scenes including default and sunset/night lighting variants), with randomized camera trajectories. Each frame ships an RGBA image, a semantic segmentation mask color-coded per object instance, and 2D tight bounding boxes.
• Real (Ours) — real photographs captured at 113 distinct physical locations covering the 10 shared object classes. Annotations were authored as polygon segmentations in COCO format and rasterized to the unified palettized PNG mask format used here.
• Real (Curated) — frames sampled from public segmentation datasets (source_dataset = "curation" or "mapillary"). Class IDs were remapped from each source taxonomy to the global BLV palette. The original per-frame split (split_origin) is preserved as a column; all curated rows are exposed under a single train split here.
• 3D Assets (synthetic_objects/) — 500 per-asset folders, each containing a .glb (Khronos), a Gaussian-splat .ply, and a .usdz (USD bundle, Apple/Pixar) ready for IsaacSim or AR pipelines. Assets are organized by BLV class.

Preprocessing

Produced by scripts/build_hf_dataset.py. Synthetic RGB PNGs are hardlinked unchanged from the source tree; the IsaacSim RGBA-encoded semantic masks are converted into single-channel palettized PNGs against a global class index; synthetic 2D bounding-box .npy files are flattened into JSONL columns; the real-world COCO polygon annotations are rasterized to the same palettized PNG format using pycocotools.

Known limitations

• Resolution mismatch. Synthetic frames are 1280×720; real-world frames are 640×360. Models that resize to a common input shape are unaffected.
• Class imbalance in real-world data. Some classes have few real-world examples (e.g. crosswalk, handrail). Report per-class mIoU alongside any aggregate.
• turnstile is synthetic-only. Evaluate over the 10 shared classes for real-world metrics.
• Sim-to-real gap. Synthetic textures and lighting may not match real-world distributions perfectly.

Ethical considerations

• The synthetic data contains no personally identifiable information.
• Real-world captures were collected in public spaces; the dataset is intended for accessibility research and must not be used for surveillance or identification of individuals.
• The class taxonomy targets infrastructure relevant to blind/low-vision navigation; models trained on this dataset should not be deployed in safety-critical settings without additional validation.

License

Released under CC BY 4.0.

Citation

@inproceedings{blv2026,
  title     = {BLV Object Recognition: A Synthetic and Real-World Benchmark},
  author    = {Anonymized Authors},
  booktitle = {NeurIPS 2026 Datasets and Benchmarks Track},
  year      = {2026}
}