Update README.md

README.md

@@ -11,6 +11,7 @@ tags:
 - fiftyone
 - group
 - image-classification
+- 3d
 dataset_summary: '
 
 
@@ -46,7 +47,7 @@ dataset_summary: '
 
   # Note: other available arguments include ''max_samples'', etc
 
-  dataset = load_from_hub("harpreetsahota/UCO3D_Dataset")
+  dataset = load_from_hub("Voxel51/UCO3D_Dataset")
 
 
   # Launch the App
@@ -60,11 +61,7 @@ dataset_summary: '
 
 # Dataset Card for UCO3D_Dataset
 
-<!-- Provide a quick summary of the dataset. -->
-
-
-
-
+[!image/png](uco3d.gif)
 
 This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 52 samples.
 
@@ -84,141 +81,171 @@ from fiftyone.utils.huggingface import load_from_hub
 
 # Load the dataset
 # Note: other available arguments include 'max_samples', etc
-dataset = load_from_hub("harpreetsahota/UCO3D_Dataset")
+dataset = load_from_hub("Voxel51/UCO3D_Dataset")
 
 # Launch the App
 session = fo.launch_app(dataset)
 ```
 
 
-## Dataset Details
+# Dataset Details
 
-### Dataset Description
+## Dataset Description
 
-<!-- Provide a longer summary of what this dataset is. -->
+uCO3D (UnCommon Objects in 3D) is a large-scale dataset of high-quality 360° videos of real-life objects with comprehensive 3D annotations. 
 
+It contains 170,000 objects across more than 1,000 diverse object categories derived from the LVIS taxonomy. Each object is captured in a turntable-like video with full 360° coverage, ensuring all sides of the object are visible. The dataset provides accurate 3D annotations including camera parameters, point clouds, and complete 3D Gaussian Splatting reconstructions. All objects are normalized to a canonical reference frame, facilitating training of generative models. 
 
+With its combination of scale, diversity, and quality, uCO3D addresses limitations in existing datasets, providing a superior resource for 3D deep learning research and applications.
 
-- **Curated by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
+- **Curated by:** Xingchen Liu, Piyush Tayal, Jianyuan Wang, Jesus Zarzar, Tom Monnier, Konstantinos Tertikas, Jiali Duan, Antoine Toisoul, Jason Y. Zhang, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, David Novotny (Meta AI and collaborating institutions)
+- **Funded by:** Meta AI
+- **Shared by:** Meta AI
 - **Language(s) (NLP):** en
-- **License:** [More Information Needed]
+- **License:** Not explicitly stated in the paper (likely Meta AI research license)
 
-### Dataset Sources [optional]
+## Dataset Sources
 
-<!-- Provide the basic links for the dataset. -->
+- **Repository:** https://github.com/facebookresearch/uco3d
+- **Hugging Face Dataset Repo:** https://huggingface.co/datasets/facebook/uco3d
+- **Paper:** arXiv:2501.07574v1 [cs.CV] (Jan 13, 2025)
+- **Demo:** https://uco3d.github.io
 
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
+# Uses
 
-## Uses
+## Direct Use
 
-<!-- Address questions around how the dataset is intended to be used. -->
+The uCO3D dataset is specifically designed for:
+- Training feedforward few-view 3D reconstruction models (like LightplaneLRM)
+- Novel view synthesis using diffusion models (CAT3D-like approaches)
+- Text-to-3D generation (Instant3D-like pipelines)
+- Training large 3D generative models
+- Digital twinning applications (creating 3D models of real-life objects)
+- Research in 3D representation learning
+- Transfer learning for 3D vision tasks
+- Benchmarking 3D reconstruction algorithms
 
-### Direct Use
+## Out-of-Scope Use
 
-<!-- This section describes suitable use cases for the dataset. -->
+The dataset is not intended for:
+- Real-time applications without appropriate downstream models
+- Human activity or behavioral analysis
+- Surveillance or security applications
+- Generating photorealistic synthetic human identities
+- Classes not represented in the 1,070 LVIS categories
+- Medical or safety-critical applications
+- Applications requiring perfect geometric accuracy (as SfM reconstructions have inherent limitations)
 
-[More Information Needed]
+# Dataset Structure
 
-### Out-of-Scope Use
+The full dataset consists of 170,000 video sequences organized by object category. Each sequence contains:
 
-<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->
+- High-resolution video (60%+ at 1080p+) with full 360° coverage of a single object
+- 200 uniformly sampled frames with corresponding camera parameters (intrinsics and extrinsics)
+- Object segmentation masks for each frame
+- Sparse and dense SfM point clouds
+- Complete 3D Gaussian Splat reconstruction
+- Textual caption describing the object
+- Normalized object-centric coordinate system alignment
 
-[More Information Needed]
+**Preview Subset:** Users can download a small 52-video subset of uCO3D for preview and debugging purposes. This subset represents a random sample of categories and takes approximately 9.6 GB, which is more than 1000x smaller than the full dataset. This preview subset is ideal for initial experimentation and workflow development before committing to the full dataset.
 
-## 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. -->
-
-[More Information Needed]
+The 1,070 object categories are organized into 50 super-categories, with approximately 20 subcategories per super-category.
 
 ## Dataset Creation
 
 ### Curation Rationale
 
-<!-- Motivation for the creation of this dataset. -->
+uCO3D was created to address limitations in existing 3D datasets:
+- CO3Dv2 has limited diversity (only 50 categories) and variable quality
+- MVImgNet lacks full 360° coverage of objects
+- High-quality 3D scan datasets like OmniObject3D have very limited scale
+- Synthetic datasets like Objaverse lack realism
 
-[More Information Needed]
+uCO3D aims to provide the optimal balance between quality, scale, and diversity for real object-centric 3D learning.
 
 ### 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. -->
-
-[More Information Needed]
+- Videos were captured by workers on Amazon Mechanical Turk
+- Workers followed a sine-wave trajectory (varying elevation) instead of a simple circular path
+- Each video was manually assessed for quality
+- Object segmentation used text-conditioned Segment-Anything (langSAM) with XMem for temporal consistency
+- 3D reconstruction used VGGSfM (more accurate than COLMAP used in previous datasets)
+- Scene alignment procedure normalized objects to canonical reference frames
+- 3D Gaussian Splat reconstruction was performed for each object using gsplat
+- Captioning used a vision-language model for per-view captions, summarized with LLAMA3
 
 #### Who are the source data producers?
 
-<!-- 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. -->
-
-[More Information Needed]
+Amazon Mechanical Turk workers captured the videos using consumer cameras. The workers were tasked with recording full 360° turntable-like videos of various objects, ensuring high resolution and complete coverage.
 
-### Annotations [optional]
-
-<!-- If the dataset contains annotations which are not part of the initial data collection, use this section to describe them. -->
+### Annotations
 
 #### Annotation process
 
-<!-- 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. -->
+- Camera parameters: VGGSfM estimated intrinsic and extrinsic parameters for 200 frames per video
+- Point clouds: VGGSfM generated sparse and dense point clouds 
+- Segmentation: langSAM provided initial segmentations, refined by XMem for temporal consistency
+- 3D reconstruction: Each scene was reconstructed using 3D Gaussian Splatting
+- Alignment: Objects were aligned to canonical reference frames through a multi-step procedure
+- Captioning: Vision-language models generated per-view captions, synthesized by LLAMA3
 
-[More Information Needed]
+The annotation quality was validated through both automated metrics and manual verification.
 
 #### Who are the annotators?
 
-<!-- This section describes the people or systems who created the annotations. -->
-
-[More Information Needed]
+A combination of:
+- Automated systems (VGGSfM, langSAM, XMem, 3D Gaussian Splatting, vision-language models)
+- Meta AI research team (for quality control and validation)
+- Amazon Mechanical Turk workers (for initial video capture)
 
 #### Personal and Sensitive Information
 
-<!-- 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. -->
-
-[More Information Needed]
+The dataset focuses on object-centric videos and does not explicitly contain personal information. The objects captured are everyday items, and care was taken during the collection process to focus solely on the objects themselves.
 
 ## Bias, Risks, and Limitations
 
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
-
-[More Information Needed]
+- Category distribution may not be perfectly balanced
+- Quality variations may exist across different object categories
+- Object selection may reflect cultural biases of the data collectors
+- SfM and 3DGS reconstructions have inherent technical limitations
+- Some complex object materials (transparent, reflective) may have lower quality reconstructions
+- The dataset is limited to static objects and does not include articulated or deformable objects
+- Training on this dataset alone may not generalize to all real-world scenarios
 
 ### 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.
+- Consider the distribution of categories when training models
+- Evaluate models on diverse test sets beyond this dataset
+- Combine with synthetic data for categories underrepresented in this dataset
+- Be aware of potential biases in object representation
+- For highest quality results, filter by reconstruction quality metrics
+- For generative models, consider augmenting with other datasets for improved diversity
 
-## Citation [optional]
+## Dataset Card Authors
 
-<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->
+The dataset card is based on information from the paper "UnCommon Objects in 3D" by Xingchen Liu et al.
 
-**BibTeX:**
+## Dataset Card Contact
 
-[More Information Needed]
+For more information, contact the authors via the GitHub repository: https://github.com/facebookresearch/uco3d
+
+# Citation
+
+```bibtex
+@inproceedings{liu24uco3d,
+    Author = {Liu, Xingchen and Tayal, Piyush and Wang, Jianyuan
+                and Zarzar, Jesus and Monnier, Tom and Tertikas, Konstantinos
+                and Duan, Jiali and Toisoul, Antoine and Zhang, Jason Y.
+                and Neverova, Natalia and Vedaldi, Andrea
+                and Shapovalov, Roman and Novotny, David},
+    Booktitle = {arXiv},
+    Title = {UnCommon Objects in 3D},
+    Year = {2024},
+}
+```
 
 **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]
+Liu, X., Tayal, P., Wang, J., Zarzar, J., Monnier, T., Tertikas, K., Duan, J., Toisoul, A., Zhang, J. Y., Neverova, N., Vedaldi, A., Shapovalov, R., & Novotny, D. (2025). UnCommon Objects in 3D. arXiv preprint arXiv:2501.07574.