README.md

metadata

annotations_creators:
  - expert-generated
language:
  - en
license: apple-ascl
multilinguality: monolingual
pretty_name: Apple Dense Material Segmentation (DMS) – Stratified Split
size_categories:
  - 10K<n<100K
source_datasets:
  - extended|open-images-v7
tags:
  - material-segmentation
  - semantic-segmentation
  - dense-prediction
  - materials
  - segformer
  - mask2former
  - stratified-split
task_categories:
  - image-segmentation
task_ids:
  - semantic-segmentation
dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype: image
    - name: image_id
      dtype: string
  splits:
    - name: train
      num_examples: 33118
    - name: validation
      num_examples: 4138
    - name: test
      num_examples: 4140

Apple Dense Material Segmentation (DMS) – Stratified 80/10/10 Split

A pixel-level material segmentation dataset containing ~41K images with dense annotations across 57 material categories. Originally released by Apple as part of the Dense Material Segmentation (DMS) research project.

This version uses a custom stratified 80/10/10 split (vs Apple's original 54/23/23) to maximise training data while maintaining representative validation and test sets.

Why a Custom Split?

Apple's original split reserves nearly half the data for evaluation (23% val + 23% test). Our re-split allocates 80% to training while using stratified sampling (based on the dominant material class per image) to keep val/test distributions aligned with the training set.

Split Quality Comparison

Metric	Original (Apple)	Custom (Stratified)	Improvement
Train size	22,492 (54%)	33,118 (80%)	+47% more training data
JSD train↔val	0.0524	0.0158	✅ 70% lower divergence
JSD train↔test	0.0526	0.0163	✅ 69% lower divergence
Classes in all splits	53/57	53/57	Equal coverage

JSD = Jensen-Shannon Divergence between pixel-level class distributions. Lower values mean the evaluation sets better represent the training distribution, leading to more reliable metrics.

Dataset Description

Each sample consists of:

Field	Type	Description
image	PIL.Image	RGB input image
label	PIL.Image	Single-channel segmentation mask (pixel values = class indices 0–56)
image_id	string	Unique image identifier

Splits

Split	Samples	Percentage
Train	33,118	80.0%
Validation	4,138	10.0%
Test	4,140	10.0%
Total	41,396	100%

Material Classes (57)

Click to expand full class list

ID	Material	ID	Material	ID	Material
0	No label	19	Gemstone/quartz	38	Sky
1	Animal skin	20	Glass	39	Snow
2	Bone/teeth/horn	21	Hair	40	Soap
3	Brickwork	22	I cannot tell	41	Soil/mud
4	Cardboard	23	Ice	42	Sponge
5	Carpet/rug	24	Leather	43	Stone, natural
6	Ceiling tile	25	Liquid, non-water	44	Stone, polished
7	Ceramic	26	Metal	45	Styrofoam
8	Chalkboard/blackboard	27	Mirror	46	Tile
9	Clutter	28	Not on list	47	Wallpaper
10	Concrete	29	Paint/plaster/enamel	48	Water
11	Cork/corkboard	30	Paper	49	Wax
12	Engineered stone	31	Pearl	50	Whiteboard
13	Fabric/cloth	32	Photograph/painting	51	Wicker
14	Fiberglass wool	33	Plastic, clear	52	Wood
15	Fire	34	Plastic, non-clear	53	Wood, tree
16	Foliage	35	Rubber/latex	54	Bad polygon
17	Food	36	Sand	55	Multiple materials
18	Fur	37	Skin/lips	56	Asphalt

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("AllanK24/apple-dms-materials-v2")

# Access splits
train_ds = dataset["train"]      # 33,118 samples
val_ds = dataset["validation"]   #  4,138 samples
test_ds = dataset["test"]        #  4,140 samples

# View a sample
sample = train_ds[0]
sample["image"].show()    # RGB image
sample["label"].show()    # Segmentation mask

Training with SegFormer / Mask2Former

from transformers import SegformerForSemanticSegmentation, SegformerImageProcessorFast
import json
from huggingface_hub import hf_hub_download

# Load class info
class_info_path = hf_hub_download(
    repo_id="AllanK24/apple-dms-materials-v2",
    filename="class_info.json",
    repo_type="dataset",
)
with open(class_info_path) as f:
    class_info = json.load(f)

id2label = {int(k): v for k, v in class_info["id2label"].items()}
label2id = class_info["label2id"]
num_labels = class_info["num_labels"]

# Initialize model
model = SegformerForSemanticSegmentation.from_pretrained(
    "nvidia/segformer-b2-finetuned-ade-512-512",
    num_labels=num_labels,
    id2label=id2label,
    label2id=label2id,
    ignore_mismatched_sizes=True,
)

# Initialize processor
processor = SegformerImageProcessorFast.from_pretrained(
    "nvidia/segformer-b2-finetuned-ade-512-512"
)

# Apply transforms
def transforms(batch):
    images = [x.convert("RGB") for x in batch["image"]]
    labels = [x for x in batch["label"]]
    return processor(images=images, segmentation_maps=labels, return_tensors="pt")

train_ds.set_transform(transforms)

Stratification Method

The split was created using a two-level stratified sampling approach:

1. Dominant class extraction – For each image, the material class with the most pixels (excluding "No label") is identified.
2. First split – Images are stratified into 80% train vs 20% eval using StratifiedShuffleSplit.
3. Second split – The 20% eval pool is stratified 50/50 into validation and test.
4. Rare class handling – Classes with <5 total images go directly to train; classes with <2 images in the eval pool are randomly assigned between val/test.

Seed: 42 (for reproducibility).

Source & Preparation

• Original dataset: Apple DMS with images from Open Images V7
• Original split (v1): AllanK24/apple-dms-materials
• Preparation pipeline: Download → resize/align (prepare_images.py) → validate (check_images.py, 41,385/41,396 passed) → stratified re-split

Citation

@article{upchurch2022dense,
  title={Dense Material Segmentation with Context-Aware Network},
  author={Upchurch, Paul and Niu, Ransen},
  year={2022},
  url={https://machinelearning.apple.com/research/dense-material-segmentation}
}

License

Released under the Apple Sample Code License (ASCL). Source images are from Open Images V7 (primarily CC BY 2.0). See the original repository for full licensing details.