Update README.md

README.md

@@ -3,24 +3,24 @@ annotations_creators: []
 language: en
 size_categories:
 - 10K<n<100K
-task_categories: []
 task_ids: []
 pretty_name: SynthHuman
 tags:
 - fiftyone
 - group
-dataset_summary: '
+dataset_summary: >
 
 
 
 
-  This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 3000 samples.
+  This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 3000
+  samples.
 
 
   ## Installation
 
 
-  If you haven''t already, install FiftyOne:
+  If you haven't already, install FiftyOne:
 
 
   ```bash
@@ -42,9 +42,9 @@ dataset_summary: '
 
   # Load the dataset
 
-  # Note: other available arguments include ''max_samples'', etc
+  # Note: other available arguments include 'max_samples', etc
 
-  dataset = load_from_hub("harpreetsahota/SynthHuman")
+  dataset = load_from_hub("Voxel51/SynthHuman")
 
 
   # Launch the App
@@ -52,8 +52,7 @@ dataset_summary: '
   session = fo.launch_app(dataset)
 
   ```
-
-  '
+license: cdla-permissive-2.0
 ---
 
 # Dataset Card for SynthHuman
@@ -82,7 +81,7 @@ from fiftyone.utils.huggingface import load_from_hub
 
 # Load the dataset
 # Note: other available arguments include 'max_samples', etc
-dataset = load_from_hub("harpreetsahota/SynthHuman")
+dataset = load_from_hub("Voxel51/SynthHuman")
 
 # Launch the App
 session = fo.launch_app(dataset)
@@ -93,130 +92,151 @@ session = fo.launch_app(dataset)
 
 ### Dataset Description
 
-<!-- Provide a longer summary of what this dataset is. -->
+The SynthHuman dataset is a high-fidelity synthetic dataset created for training human-centric computer vision models. It contains 300,000 high-resolution (384×512) images with three main tasks: relative depth estimation, surface normal estimation, and soft foreground segmentation. The dataset features procedurally-generated human subjects in diverse poses, environments, lighting, and appearances, with equal distribution of faces, upper body, and full body scenarios. 
+
+Unlike scan-based synthetic datasets, SynthHuman uses high-fidelity procedural generation techniques to create detailed human representations, including strand-level hair (with hundreds of thousands of individual 3D strands per hairstyle), detailed clothing, accessories, and expressive faces. This approach enables the generation of ground-truth annotations with strand-level granularity that capture fine details like facial wrinkles, eyelids, hair strands, and subtle texture variations.
 
+- **Curated by:** Microsoft Research, Cambridge, UK
 
+- **Funded by:** Microsoft
+
+- **Shared by:** Microsoft
 
-- **Curated by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
 - **Language(s) (NLP):** en
-- **License:** [More Information Needed]
 
-### Dataset Sources [optional]
+- **License:** [CDLA - Permissive - 2.0](https://github.com/microsoft/DAViD/blob/main/LICENSE-CDLA-2.0.txt)
 
-<!-- Provide the basic links for the dataset. -->
+### Dataset Sources
 
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
+- **Repository:** https://aka.ms/DAViD
+  
+- **Paper:** DAViD: Data-efficient and Accurate Vision Models from Synthetic Data (arXiv:2507.15365)
 
-## Uses
+- **Parsing to FiftyOne format:** https://github.com/harpreetsahota204/synthhuman_to_fiftyone
 
-<!-- Address questions around how the dataset is intended to be used. -->
+## Uses
 
 ### Direct Use
 
-<!-- This section describes suitable use cases for the dataset. -->
+The SynthHuman dataset is designed for training computer vision models for human-centric dense prediction tasks, specifically:
 
-[More Information Needed]
+1. **Relative depth estimation**: Predicting per-pixel depth values for human subjects
 
-### Out-of-Scope Use
+2. **Surface normal estimation**: Predicting per-pixel surface normal vectors (xyz components)
+ 
+3. **Soft foreground segmentation**: Generating soft alpha masks to separate humans from backgrounds
 
-<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->
+The dataset enables training smaller, more efficient models that achieve state-of-the-art accuracy without requiring large-scale pretraining or complex multi-stage training pipelines. This makes it suitable for applications with computational constraints.
+
+### Out-of-Scope Use
 
-[More Information Needed]
+The dataset should not be used for:
+- Identifying or recognizing specific individuals
+- Creating deceptive or misleading synthetic human content
+- Applications that could violate privacy or cause harm to real individuals
+- Training models for tasks beyond the three specified dense prediction tasks without proper evaluation
 
 ## Dataset Structure
 
-<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->
+The SynthHuman dataset contains 300,000 synthetic images of resolution 384×512, with equal distribution (100,000 each) across three categories:
+1. Face scenarios
+2. Upper body scenarios
+3. Full body scenarios
 
-[More Information Needed]
+Each sample in the dataset includes:
+- RGB rendered image
+- Soft foreground mask (alpha channel)
+- Surface normals (3-channel)
+- Depth ground-truth annotations
+
+The dataset is designed to be diverse in terms of:
+- Human poses and expressions
+- Environments and lighting conditions
+- Physical appearances (body shapes, clothing, accessories)
+- Camera viewpoints
 
 ## Dataset Creation
 
 ### Curation Rationale
 
-<!-- Motivation for the creation of this dataset. -->
-
-[More Information Needed]
+The dataset was created to address limitations in existing human-centric computer vision datasets, which often suffer from:
+1. Imperfect ground truth annotations due to reliance on photogrammetry or noisy sensors
+2. Limited diversity in subjects and environments due to challenges in capturing in-the-wild data
+3. Inability to capture fine details like hair strands, reflective surfaces, and subtle geometric features
 
 ### Source Data
 
-<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->
-
 #### Data Collection and Processing
 
-<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->
+The dataset generation process involved sampling from:
 
-[More Information Needed]
+- Face/body shapes (from training sources and a library of 3572 scans)
+  
+- Expressions and poses (from AMASS, MANO, and other sources)
 
-#### Who are the source data producers?
+- Textures (from high-resolution face scans with expression-based dynamic wrinkle maps)
 
-<!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. -->
+- Hair styles (548 strand-level 3D hair models, each with 100K+ strands)
 
-[More Information Needed]
+- Accessories (36 glasses, 57 headwear items)
 
-### Annotations [optional]
+- Clothing (50+ clothing tops)
 
-<!-- If the dataset contains annotations which are not part of the initial data collection, use this section to describe them. -->
+- Environments (mix of HDRIs and 3D environments)
 
-#### Annotation process
+Rendering the complete dataset took 72 hours on a cluster of 300 machines with M60 GPUs (equivalent to 2 weeks on a 4-GPU A100 machine).
 
-<!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->
+#### Who are the source data producers?
 
-[More Information Needed]
+The dataset was created by researchers at Microsoft Research in Cambridge, UK. The synthetic data was procedurally generated using artist-created assets, scanned data sources, and procedural generation techniques.
 
-#### Who are the annotators?
+### Annotations
 
-<!-- This section describes the people or systems who created the annotations. -->
+#### Annotation process
 
-[More Information Needed]
+Since this is a synthetic dataset, the annotations are generated programmatically during the rendering process rather than being manually created. This ensures perfect alignment between the RGB images and their corresponding ground truths.
 
-#### Personal and Sensitive Information
+Special attention was given to:
+1. **Hair representation**: A voxel-grid volume with density based on strand geometry was created, then converted to a coarse proxy mesh using marching cubes to generate interpretable normal vectors.
+2. **Transparent surfaces**: The dataset provides control over whether depth and normals of translucent surfaces (like eyeglass lenses) are included or whether they show the surface behind them.
+3. **Soft foreground masks**: Generated with pixel-perfect accuracy, including partial transparency for hair strands and other fine structures.
 
-<!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->
+#### Personal and Sensitive Information
 
-[More Information Needed]
+The dataset contains only synthetic human representations and does not include any real personal or sensitive information. The synthetic data generation process ensures that no real individuals are represented in the dataset.
 
 ## Bias, Risks, and Limitations
 
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
-
-[More Information Needed]
+While the paper doesn't explicitly discuss biases in the dataset, there are potential limitations:
+- The paper notes that there may be aspects of human diversity not yet represented in the dataset
+- The synthetic nature of the data might not fully capture all real-world scenarios and edge cases
+- Models trained on this data may have lower accuracy for some demographic groups (acknowledged as a potential issue in the paper)
+- Failure cases noted in the paper include extreme lighting conditions, printed patterns on clothing, tattoos, and rare scale variations (e.g., a baby held in an adult's hand)
 
 ### Recommendations
 
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
-
-Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.
-
-## Citation [optional]
+Users should be aware of the following:
+- The dataset creators acknowledge that the synthetic data approach helps address fairness concerns by providing precise control over the training data distribution
+- Additional diversity in assets and scene variations could improve robustness to real-world scenarios
+- Users should test models trained on this data across diverse real-world populations to ensure fair performance
+- For applications involving human subjects, users should consider the ethical implications and potential biases
+- Supplementing with real-world data for specific challenging scenarios might be beneficial
 
-<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->
+## Citation
 
 **BibTeX:**
 
-[More Information Needed]
+```bibtex
+@misc{saleh2025david,
+    title={{DAViD}: Data-efficient and Accurate Vision Models from Synthetic Data},
+    author={Fatemeh Saleh and Sadegh Aliakbarian and Charlie Hewitt and Lohit Petikam and Xiao-Xian and Antonio Criminisi and Thomas J. Cashman and Tadas Baltrušaitis},
+    year={2025},
+    eprint={2507.15365},
+    archivePrefix={arXiv},
+    primaryClass={cs.CV},
+    url={https://arxiv.org/abs/2507.15365},
+}
+```
 
 **APA:**
-
-[More Information Needed]
-
-## Glossary [optional]
-
-<!-- If relevant, include terms and calculations in this section that can help readers understand the dataset or dataset card. -->
-
-[More Information Needed]
-
-## More Information [optional]
-
-[More Information Needed]
-
-## Dataset Card Authors [optional]
-
-[More Information Needed]
-
-## Dataset Card Contact
-
-[More Information Needed]
+Saleh, F., Aliakbarian, S., Hewitt, C., Petikam, L., Criminisi, A., Cashman, T. J., & Baltrusaitis, T. (2025). DAViD: Data-efficient and Accurate Vision Models from Synthetic Data. arXiv preprint arXiv:2507.15365.