GSD-S: Glass Surface Detection – Semantics

GSD-S is a glass surface detection dataset augmented with per-pixel semantic labels, introduced in the NeurIPS 2022 paper "Exploiting Semantic Relations for Glass Surface Detection". Each sample pairs an RGB photograph with a binary glass mask and a 43-class semantic segmentation map, enabling joint glass detection and scene-semantic reasoning.

• Paper: Exploiting Semantic Relations for Glass Surface Detection — NeurIPS 2022
• Project page: https://jiaying.link/neurips2022-gsds/
• Authors: Jiaying Lin, Yuen-Hei Yeung, Rynson W.H. Lau (City University of Hong Kong)

---

Dataset Summary

Split	Samples
train	3,911
test	608
total	4,519

Images are 640 × 480 pixels (JPEG). All annotation maps are PNG.

---

Columns

Column	Type	Description
image_id	string	Original filename stem (e.g. 000000000711); use for round-trip fidelity
image	Image	RGB photograph (.jpg)
mask	Image	Binary glass mask — pixel values 0 (non-glass) or 255 (glass)
seg	Image	Semantic segmentation map — pixel values 0–42 (class index)
seg_colored	Image	False-color rendering of seg using the GSD-S palette (for visualization)

Semantic classes (43 total)

unknown, wall, glass, floor, ceiling, door, chair, table, sofa, cabinet, curtain, blinds, bedding, picture, light, clothes, counter, sink, toilet, towel, mirror, tv, building_structure, stationery, plant, person, fridge, bath_shower, seat, floor_mat, fence, ground, bottle, kitchenware, road, transport, electronics, food, bag, nature, animal, road_infrastructure, clock

The class-to-color mapping is available in the official repository at utils/GSD-S_color_map.csv.

---

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("garrying/GSD-S")
sample = ds["train"][0]

print(sample["image_id"])   # e.g. "000000000711"
sample["image"].show()      # RGB photo
sample["mask"].show()       # binary glass mask
sample["seg"].show()        # semantic class indices
sample["seg_colored"].show() # false-color visualization

---

Converting Back to Raw Files

A conversion helper is bundled in this repository. Download and run it:

# Download the script
huggingface-cli download garrying/GSD-S parquet_to_raw.py --repo-type dataset --local-dir .

# Restore all splits to ./GSD-S/
python parquet_to_raw.py --repo garrying/GSD-S

# Or restore from a locally cached copy
python parquet_to_raw.py --local /path/to/local/cache

Output layout:

GSD-S/
  train/
    images/         # .jpg
    masks/          # .png
    segs/           # .png  (class-index maps)
    segs_colored/   # .png  (false-color maps)
  test/
    ...

---

Evaluation Metrics

The official evaluation protocol reports:

• IoU — Intersection over Union
• F-measure (Fβ, β² = 0.3) — weighted precision-recall
• MAE — Mean Absolute Error
• BER — Balanced Error Rate

Predictions and ground-truth masks are binarized at threshold 0.5 before computing all metrics.

---

Citation

@inproceedings{neurips2022:gsds2022,
  title     = {Exploiting Semantic Relations for Glass Surface Detection},
  author    = {Lin, Jiaying and Yeung, Yuen Hei and Lau, Rynson W.H.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2022}
}

---

License

BSD 3-Clause License — non-commercial use only. See LICENSE for the full text. Please cite the paper if you use this dataset.