Dataset Viewer
Auto-converted to Parquet Duplicate
image
image
label
image
image_id
string
56515
37443
23443
42893
23551
29991
213305
39401
28439
32588
41732
38972
197427
31884
57923
203717
39905
41681
33958
126469
29836
28425
60777
45710
34508
40997
38471
220270
42351
33401
60405
60106
56178
45996
34352
127167
25283
35601
26227
58174
91982
209585
27172
190132
51653
217303
35407
43517
208171
33801
33297
215786
31631
38631
33506
26466
44132
60919
35498
209104
29184
51462
58368
60274
50448
24596
24481
149200
24237
54915
44234
35264
37519
42218
27149
193358
26932
214899
23530
46373
55414
198816
50615
35386
33846
24968
42708
54628
29256
42302
41119
44044
207266
209012
35864
28665
56061
130478
32381
58916
End of preview. Expand in Data Studio

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.

Downloads last month
17