Upload README.md with huggingface_hub

README.md

@@ -1,4 +1,26 @@
 ---
+annotations_creators:
+- expert-generated
+language:
+- en
+license: apple-ascl
+multilinguality: monolingual
+pretty_name: Apple Dense Material Segmentation (DMS)
+size_categories:
+- 10K<n<100K
+source_datasets:
+- extended|open-images-v7
+tags:
+- material-segmentation
+- semantic-segmentation
+- dense-prediction
+- materials
+- segformer
+- mask2former
+task_categories:
+- image-segmentation
+task_ids:
+- semantic-segmentation
 dataset_info:
   features:
   - name: image
@@ -9,126 +31,158 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 6251674501
     num_examples: 22492
   - name: validation
-    num_bytes: 2584262537
     num_examples: 9412
   - name: test
-    num_bytes: 2581775314
     num_examples: 9492
-  download_size: 11729195090
-  dataset_size: 11417712352
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: validation
-    path: data/validation-*
-  - split: test
-    path: data/test-*
 ---
 
+# Apple Dense Material Segmentation (DMS) Dataset
 
-```markdown
----
-license: cc-by-nc-4.0
-task_categories:
-- image-segmentation
-- semantic-segmentation
-language:
-- en
-tags:
-- material-segmentation
-- apple-dms
-- scene-parsing
-- segformer
-size_categories:
-- 10k<n<100k
----
+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)](https://machinelearning.apple.com/research/dense-material-segmentation) research project.
 
-# Apple Dense Material Segmentation (DMS) Dataset
+> **Note**: This is a mirror prepared for direct use with the HuggingFace 🤗 `datasets` library. The source images originate from [Open Images V7](https://storage.googleapis.com/openimages/web/index.html), and material annotations were created by Apple. Some images (~6%) from the original dataset could not be retrieved from Open Images and are therefore absent.
+
+## 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
 
-This dataset is a reconstructed version of the **Apple Dense Material Segmentation (DMS)** dataset, originally introduced in the ECCV 2022 paper *"A Dense Material Segmentation Dataset for Indoor and Outdoor Scene Parsing"*.
+| Split | Samples | Percentage |
+|-------|---------|------------|
+| Train | 22,492 | 54.3% |
+| Validation | 9,412 | 22.7% |
+| Test | 9,492 | 22.9% |
+| **Total** | **41,396** | **100%** |
 
-It pairs high-quality dense material masks (provided by Apple) with the corresponding real-world images from the Open Images dataset.
+The split assignments follow the original Apple DMS partition.
 
-**Repository:** [AllanK24/apple-dms-materials](https://huggingface.co/datasets/AllanK24/apple-dms-materials)
+### Material Classes (57)
 
-## Dataset Summary
+<details>
+<summary>Click to expand full class list</summary>
 
-- **Total Samples:** ~41,385 image-mask pairs
-- **Classes:** 46 material categories (e.g., Wood, Glass, Plastic, Metal, Fabric)
-- **Task:** Semantic Segmentation
-- **Origin:** Reassembled from [Apple's Official Repository](https://github.com/apple/ml-dms-dataset) and Open Images archives.
+| 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 |
 
-### Missing Data Note
-This dataset was reconstructed in 2026. Due to "internet rot" (images being deleted from Flickr and Open Images S3 buckets over the years), approximately **~6% of the original images are permanently lost**. 
-- **Original Dataset Size:** ~44,500 images
-- **Current Dataset Size:** ~41,385 images
-- **Status:** The dataset is ~94% complete and fully functional for training state-of-the-art segmentation models.
+</details>
 
 ## Usage
 
-You can load this dataset directly using the Hugging Face `datasets` library.
+### Loading the Dataset
 
 ```python
 from datasets import load_dataset
 
-# Load the dataset
 dataset = load_dataset("AllanK24/apple-dms-materials")
 
-# Inspect a single example
-example = dataset["train"][0]
-image = example["image"]
-mask = example["label"]
-image_id = example["image_id"]
-
-print(f"Loaded image {image_id} with size {image.size}")
+# Access splits
+train_ds = dataset["train"]
+val_ds = dataset["validation"]
+test_ds = dataset["test"]
 
+# View a sample
+sample = train_ds[0]
+print(sample["image_id"])   # e.g. "22491"
+sample["image"].show()      # RGB image
+sample["label"].show()      # Segmentation mask
 ```
 
-### Class Labels & Taxonomy
-
-The dataset includes a `class_info.json` file containing the mapping between label IDs and material names.
-
-* **Background/Void:** Depending on the configuration, ID 0 or 255 is often used for void. Check `class_info.json` for the exact ID mapping.
-* **Classes:** 46 standard materials.
+### Training with SegFormer
 
-## Dataset Structure
+```python
+from transformers import SegformerForSemanticSegmentation, SegformerImageProcessorFast
+import json
+
+# Load class info
+from huggingface_hub import hf_hub_download
+class_info_path = hf_hub_download(
+    repo_id="AllanK24/apple-dms-materials",
+    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"
+)
+```
 
-The dataset is divided into three standard splits:
+### Applying Transforms
 
-| Split | Samples |
-| --- | --- |
-| `train` | ~30,000 |
-| `validation` | ~5,000 |
-| `test` | ~6,000 |
+```python
+def transforms(batch):
+    images = [x.convert("RGB") for x in batch["image"]]
+    labels = [x for x in batch["label"]]
+    inputs = processor(images=images, segmentation_maps=labels, return_tensors="pt")
+    return inputs
 
-## Licensing
+train_ds.set_transform(transforms)
+```
 
-This dataset is a composite of two sources with different licenses:
+## Dataset Preparation
 
-1. **Annotations (Masks):** Created by Apple. Distributed under the license specified in the [Apple ML-DMS Repository](https://github.com/apple/ml-dms-dataset) (Copyright © 2022 Apple Inc.).
-2. **Images:** Sourced from Open Images (Flickr). These images retain their original Creative Commons licenses (typically CC-BY 2.0 or CC-BY 4.0) as defined by their original photographers.
+This dataset was prepared from the original Apple DMS release using the following pipeline:
 
-**Disclaimer:** This repository is a redistribution of the reconstructed dataset for research convenience. Users are responsible for adhering to the licenses of both the image sources and the annotation providers.
+1. **Download** – Source images retrieved from Open Images V7 using URLs in Apple's metadata.
+2. **Resize & align** – Images resized to match label dimensions using Apple's [`prepare_images.py`](https://github.com/apple/ml-dms-dataset).
+3. **Validation** – Image–label consistency verified with Apple's `check_images.py` (41,385 / 41,396 passed; 11 minor rotation warnings).
 
 ## Citation
 
 If you use this dataset, please cite the original Apple paper:
 
 ```bibtex
-@inproceedings{upchurch2022dense,
-  title={A Dense Material Segmentation Dataset for Indoor and Outdoor Scene Parsing},
+@article{upchurch2022dense,
+  title={Dense Material Segmentation with Context-Aware Network},
   author={Upchurch, Paul and Niu, Ransen},
-  booktitle={European Conference on Computer Vision (ECCV)},
-  year={2022}
+  year={2022},
+  url={https://machinelearning.apple.com/research/dense-material-segmentation}
 }
-
 ```
 
-```
+## License
 
-```
+This dataset is released under the [Apple Sample Code License (ASCL)](https://developer.apple.com/sample-code/license/apple-sample-code-license/). The source images are from Open Images V7 and are subject to their respective licenses (primarily CC BY 2.0). Please refer to the [original repository](https://github.com/apple/ml-dms-dataset) for full licensing details.