Furqan7007/IDDAW_OFFICIAL

IDD-AW: India Driving Dataset – Adverse Weather

Semantic segmentation benchmark for autonomous driving in rain, fog, low-light, and snow, with paired RGB + near-infrared (NIR) frames and dense Level-3 semantic labels (26 classes).

TODO before publishing: confirm and set the correct license / citation for the original IDD-AW release (see iddaw.github.io and the WACV 2024 paper "IDD-AW: A Benchmark for Safe Semantic Segmentation in Adverse Weather"). This card currently marks the license as other.

Contents

Split	Drives	Frames
train	143	3430
val	18	475

Each frame has three co-registered PNGs: RGB, NIR, and a Level-3 label map. (The dataset was uploaded with one train frame — drive 175, frame 00000028 — removed because it had no ground-truth label, so every frame here is fully paired.)

Directory layout

train/ , val/
  rgb/<driveId>/<frame>_rgb.png                    # RGB image
  nir/<driveId>/<frame>_nir.png                    # near-infrared, same frame
  gt_labels/<driveId>/<frame>_labellevel3Ids.png   # Level-3 semantic label ids

• <driveId> is an integer folder (e.g. 58, 175).
• <frame> is a zero-padded 8-digit index (e.g. 00000028).
• Image suffix: _rgb.png • label suffix: _labellevel3Ids.png.
• Label maps are single-channel; pixel value = class id (see table below).

Classes (Level-3, 26 classes)

id	name	RGB palette
0	road	(128, 64, 128)
1	drivable fallback	(81, 0, 81)
2	sidewalk	(244, 35, 232)
3	non drivable fallback	(152, 251, 152)
4	person	(220, 20, 60)
5	rider	(255, 0, 0)
6	motorcycle	(0, 0, 230)
7	bicycle	(119, 11, 32)
8	autorickshaw	(255, 204, 54)
9	car	(0, 0, 142)
10	truck	(0, 0, 70)
11	bus	(0, 60, 100)
12	vehicle fallback	(0, 0, 90)
13	curb	(196, 196, 196)
14	wall	(102, 102, 156)
15	fence	(190, 153, 153)
16	guard rail	(180, 165, 180)
17	billboard	(174, 64, 67)
18	traffic sign	(220, 220, 0)
19	traffic light	(250, 170, 30)
20	pole	(153, 153, 153)
21	obs-str-bar-fallback	(169, 187, 214)
22	building	(70, 70, 70)
23	bridge	(150, 100, 100)
24	vegetation	(107, 142, 35)
25	sky	(70, 130, 180)

Download

from huggingface_hub import snapshot_download

snapshot_download(
    "Furqan7007/IDDAW_OFFICIAL",
    repo_type="dataset",
    local_dir="IDDAW",
)

Grab just one split, or one drive:

hf download Furqan7007/IDDAW_OFFICIAL --repo-type=dataset \
  --include="val/**" --local-dir IDDAW

Load a frame

from pathlib import Path
import numpy as np
from PIL import Image

root  = Path("IDDAW/train")
rgb   = np.array(Image.open(root / "rgb/58/00000000_rgb.png"))
nir   = np.array(Image.open(root / "nir/58/00000000_nir.png"))
label = np.array(Image.open(root / "gt_labels/58/00000000_labellevel3Ids.png"))

Citation

@inproceedings{iddaw2024,
  title     = {IDD-AW: A Benchmark for Safe Semantic Segmentation in Adverse Weather},
  author    = {Shaik, Furqan Ahmed and others},
  booktitle = {WACV},
  year      = {2024}
}