Update README.md

README.md

@@ -46,7 +46,7 @@ dataset_summary: '
 
   # Note: other available arguments include ''max_samples'', etc
 
-  dataset = load_from_hub("harpreetsahota/sku110k_test")
+  dataset = load_from_hub("Voxel51/sku110k_test")
 
 
   # Launch the App
@@ -60,10 +60,7 @@ dataset_summary: '
 
 # Dataset Card for harpreetsahota/sku110k_test
 
-<!-- Provide a quick summary of the dataset. -->
-
-
-
+![image](sku110k.gif)
 
 
 This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 2936 samples.
@@ -84,141 +81,211 @@ from fiftyone.utils.huggingface import load_from_hub
 
 # Load the dataset
 # Note: other available arguments include 'max_samples', etc
-dataset = load_from_hub("harpreetsahota/sku110k_test")
+dataset = load_from_hub("Voxel51/sku110k_test")
 
 # Launch the App
 session = fo.launch_app(dataset)
 ```
 
 
+# Dataset Card for SKU-110K (test split)
+
 ## Dataset Details
 
 ### Dataset Description
 
-<!-- Provide a longer summary of what this dataset is. -->
+The SKU-110K dataset is a large-scale benchmark for object detection in densely packed retail scenes. It consists of 11,762 images of retail shelves from thousands of supermarkets worldwide, encompassing diverse geographic locations including the United States, Europe, and East Asia. The dataset contains over 1.73 million bounding box annotations, with an average of approximately 147 objects per image. All images have been resized to a resolution of one million pixels.
 
+The dataset addresses the challenge of precise detection in densely packed scenes where objects are closely positioned, often overlapping, and typically oriented within a range of [-15°, 15°]. This makes it particularly valuable for developing and evaluating object detection algorithms for real-world retail applications where traditional detection methods often struggle due to extreme object density and occlusion.
 
+- **Curated by:** Eran Goldman, Roei Herzig, Aviv Eisenschtat, Jacob Goldberger, and Tal Hassner
+- **Funded by:** Trax (based on license information)
+- **Shared by:** Research team from Bar-Ilan University and Trax
+- **Language(s) (NLP):** Not applicable (computer vision dataset)
+- **License:** Academic and non-commercial use only (proprietary license by Trax)
 
-- **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
 
-### Dataset Sources [optional]
-
-<!-- Provide the basic links for the dataset. -->
-
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
+- **Repository:** https://github.com/eg4000/SKU110K_CVPR19
+- **Paper:** "Precise Detection in Densely Packed Scenes" - CVPR 2019
+- **ArXiv:** https://arxiv.org/abs/1904.00853
 
 ## Uses
 
-<!-- Address questions around how the dataset is intended to be used. -->
-
 ### Direct Use
 
-<!-- This section describes suitable use cases for the dataset. -->
+The SKU-110K dataset is designed for the following use cases:
 
-[More Information Needed]
+- **Object Detection Research:** Training and evaluating object detection models, particularly for densely packed scenes
+- **Retail Analytics:** Developing algorithms for automated shelf monitoring, inventory management, and planogram compliance
+- **Benchmark Evaluation:** Comparing performance of detection algorithms in challenging, high-density scenarios
+- **Dense Object Detection:** Research on handling extreme object density, occlusion, and scale variation
+- **Academic Research:** Educational purposes and non-commercial research projects
 
-### Out-of-Scope Use
-
-<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->
-
-[More Information Needed]
+The dataset is particularly suitable for:
+- Studying detection performance in scenes with 50-200+ objects per image
+- Developing algorithms robust to varying lighting conditions, viewpoints, and scales
+- Research on handling closely packed objects with minimal spacing
 
 ## 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 dataset is organized into three splits with CSV annotation files:
 
-[More Information Needed]
+### Split Statistics
 
-## Dataset Creation
+| Split | Images | Annotations | Avg. Objects/Image |
+|-------|--------|-------------|-------------------|
+| Train | 8,233 | 1,208,482 | ~147 |
+| Validation | 588 | 90,968 | ~155 |
+| Test | 2,941 | 431,546 | ~147 |
+| **Total** | **11,762** | **1,730,996** | **~147** |
 
-### Curation Rationale
+### Annotation Format
 
-<!-- Motivation for the creation of this dataset. -->
+The CSV annotation files contain the following columns:
 
-[More Information Needed]
+- `image_name`: Filename of the image (e.g., "test_0.jpg")
+- `x1`: X-coordinate of the top-left corner of the bounding box (pixels)
+- `y1`: Y-coordinate of the top-left corner of the bounding box (pixels)
+- `x2`: X-coordinate of the bottom-right corner of the bounding box (pixels)
+- `y2`: Y-coordinate of the bottom-right corner of the bounding box (pixels)
+- `class`: Class label (all objects labeled as "object" - no fine-grained categories)
+- `image_width`: Width of the image in pixels
+- `image_height`: Height of the image in pixels
 
-### Source Data
+**Note:** Each annotation appears on a separate line in the CSV file, meaning images with multiple objects have multiple rows.
 
-<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->
+### FiftyOne Dataset Structure
 
-#### Data Collection and Processing
+The dataset has been converted to FiftyOne format with the following enhancements:
 
-<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->
+#### Base Structure
+- **Dataset Name:** `sku110k_test` (test split)
+- **Sample Structure:** Each sample represents one image with associated detections
+- **Image Path:** `SKU110K_fixed/images/{image_name}`
+- **Detection Field:** `ground_truth` (FiftyOne Detections object)
 
-[More Information Needed]
+#### Bounding Box Format
+Bounding boxes are stored in FiftyOne's normalized format:
+- `[x, y, width, height]` where all values are in range [0, 1]
+- `x`: Normalized x-coordinate of top-left corner (x1 / image_width)
+- `y`: Normalized y-coordinate of top-left corner (y1 / image_height)
+- `width`: Normalized width ((x2 - x1) / image_width)
+- `height`: Normalized height ((y2 - y1) / image_height)
 
-#### Who are the source data producers?
+#### Enriched Fields
 
-<!-- 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. -->
+The FiftyOne dataset includes the following enrichments:
 
-[More Information Needed]
+1. **Bounding Box Areas** (`area` field on each detection)
+   - Computed as: `width × height` (in normalized coordinates)
+   - Range: [0, 1] representing the proportion of image covered
 
-### Annotations [optional]
+2. **Detection Counts** (`num_detections` field at sample level)
+   - Integer count of objects detected in each image
+   - Useful for filtering and analyzing image complexity
 
-<!-- If the dataset contains annotations which are not part of the initial data collection, use this section to describe them. -->
+3. **RADIO Embeddings** (`radio_embeddings` field at sample level)
+   - Global semantic features extracted using C-RADIO v3-h model
+   - High-dimensional vectors capturing visual semantics
+   - Enables similarity search and clustering
 
-#### Annotation process
+4. **UMAP Visualization** (Brain key: `radio_viz`)
+   - 2D projection of RADIO embeddings for visualization
+   - Allows exploration of visual similarity patterns
+   - Interactive visualization in FiftyOne App
 
-<!-- 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. -->
+5. **Attention Heatmaps** (`radio_heatmap` field at sample level)
+   - Spatial attention maps from C-RADIO v3-h model
+   - Generated with smoothing (sigma=0.51)
+   - Format: NCHW (channels first)
+   - Highlights salient regions in each image
 
-[More Information Needed]
+## Dataset Creation
 
-#### Who are the annotators?
+### Curation Rationale
 
-<!-- This section describes the people or systems who created the annotations. -->
+The SKU-110K dataset was created to address a critical gap in object detection research: the lack of large-scale datasets for densely packed scenes. While existing datasets like COCO and Pascal VOC contain object detection annotations, they typically feature relatively sparse scenes with well-separated objects. Real-world retail environments present unique challenges:
 
-[More Information Needed]
+- **Extreme Density:** Shelves contain 50-200+ products in close proximity
+- **Heavy Occlusion:** Objects frequently overlap and obscure one another
+- **Scale Variation:** Products vary greatly in size within the same scene
+- **Orientation Patterns:** Most objects aligned within [-15°, 15°] range
 
-#### Personal and Sensitive Information
+The dataset enables research on precise localization and detection algorithms capable of handling these challenging conditions, with applications in automated retail analytics, inventory management, and planogram compliance.
 
-<!-- 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. -->
+### Source Data
 
-[More Information Needed]
+#### Data Collection and Processing
 
-## Bias, Risks, and Limitations
+- **Collection Method:** Images captured from thousands of supermarket stores worldwide
+- **Geographic Diversity:** United States, Europe, and East Asia
+- **Scene Variation:** Diverse scales, viewpoints, lighting conditions, and noise levels
+- **Image Processing:** All images resized to one million pixels for consistency
+- **Quality Control:** Images selected to represent challenging, densely packed scenarios
+- **Annotation Tool:** Manual annotation using bounding box annotation software
+- **Format:** CSV files with one annotation per line
 
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+The dataset focuses on "in-the-wild" conditions with natural variations in:
+- Camera angles and distances
+- Lighting (fluorescent, natural, mixed)
+- Shelf arrangements and product placement
+- Image quality and noise levels
 
-[More Information Needed]
+#### Who are the source data producers?
+
+The source images were captured from retail stores operated by various supermarket chains across multiple continents. The images represent real retail environments and were collected through Trax, a retail technology company specializing in computer vision solutions for in-store execution.
 
-### Recommendations
+### Annotations
 
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+#### Annotation process
 
-Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.
+- **Annotation Type:** Manual bounding box annotation
+- **Annotation Guidelines:** Annotators were instructed to draw tight bounding boxes around each visible product on retail shelves
+- **Class Labels:** All objects labeled uniformly as "object" (no product-level categorization)
+- **Annotation Density:** Average of 147 bounding boxes per image, with some images containing 200+ annotations
+- **Quality Assurance:** Manual review and validation process to ensure annotation accuracy
+- **Tools Used:** Professional annotation tools for computer vision tasks
+- **Completeness:** All visible products in each image were annotated
 
-## Citation [optional]
+**Note:** The dataset does not include fine-grained product categories or SKU-level identification. All objects are labeled with a single "object" class, making this a class-agnostic detection task focused on localization precision rather than classification.
 
-<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->
+#### Who are the annotators?
 
-**BibTeX:**
+The annotations were created by trained professional annotators working with the research team. Specific demographic information about the annotators is not publicly available. The annotation process was conducted with quality control measures to ensure consistency and accuracy across the large annotation volume (1.7M+ bounding boxes).
 
-[More Information Needed]
+#### Personal and Sensitive Information
 
-**APA:**
+The dataset consists of images of retail shelf scenes containing packaged products. The images do not intentionally capture or focus on people. However, users should be aware that:
 
-[More Information Needed]
+- Retail environments are public spaces where incidental capture of individuals may occur
+- Product brands and packaging visible in images are proprietary to their respective manufacturers
+- Store layouts and product arrangements may be considered proprietary information
 
-## Glossary [optional]
+The dataset is provided with restrictions on redistribution and commercial use to protect potential proprietary interests.
 
-<!-- If relevant, include terms and calculations in this section that can help readers understand the dataset or dataset card. -->
+## Citation
 
-[More Information Needed]
+### BibTeX
 
-## More Information [optional]
+```bibtex
+@inproceedings{goldman2019dense,
+  author    = {Eran Goldman and Roei Herzig and Aviv Eisenschtat and Jacob Goldberger and Tal Hassner},
+  title     = {Precise Detection in Densely Packed Scenes},
+  booktitle = {Proc. Conf. Comput. Vision Pattern Recognition (CVPR)},
+  year      = {2019}
+}
+```
 
-[More Information Needed]
+### APA
 
-## Dataset Card Authors [optional]
+Goldman, E., Herzig, R., Eisenschtat, A., Goldberger, J., & Hassner, T. (2019). Precise Detection in Densely Packed Scenes. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)* (pp. 5227-5236).
 
-[More Information Needed]
+## More Information
 
-## Dataset Card Contact
+### Additional Resources
 
-[More Information Needed]
+- **GitHub Repository:** https://github.com/eg4000/SKU110K_CVPR19
+- **ArXiv Paper:** https://arxiv.org/abs/1904.00853
+- **FiftyOne Documentation:** https://docs.voxel51.com/
+- **RADIO Model:** https://github.com/harpreetsahota204/NVLabs_CRADIOV3