training/README.md

NVR Entity Recognition Dataset (Standardized)

Dataset Structure

This dataset follows the COCO format for object detection training:

dataset/
├── images/
│   ├── train/          # Training images (68 images)
│   ├── val/            # Validation images (19 images)
│   └── test/           # Test images (11 images)
├── annotations/
│   ├── train.json      # Training annotations (COCO format)
│   ├── val.json        # Validation annotations (COCO format)
│   └── test.json       # Test annotations (COCO format)
└── dataset_config.json # Dataset configuration and metadata

Categories

19 categories identified from folder structure:

• back-sleeping (ID: 1)
• bedroom (ID: 2)
• blanket-in-bassinet (ID: 3)
• face-down (ID: 4)
• pacifier_pacifier-fell-out (ID: 5)
• pacifier_pacifier-in-mouth (ID: 6)
• reference_subject (ID: 7)
• side-sleeping (ID: 8)
• smothering_baby-under-blanket (ID: 9)
• smothering_smothering (ID: 10)
• smothering_smothering-high-risk (ID: 11)
• subject-obscured (ID: 12)
• swaddling_normal-swaddle (ID: 13)
• swaddling_swaddle-breakout (ID: 14)
• unsafe-object-presence_button-battery (ID: 15)
• unsafe-object-presence_pill (ID: 16)
• unsafe-object-presence_screw (ID: 17)
• unsafe-object-presence_sim-opener (ID: 18)
• zone-empty (ID: 19)

Usage

This dataset can be loaded directly with popular ML frameworks:

PyTorch/Detectron2

from detectron2.data.datasets import register_coco_instances

register_coco_instances("nvr_train", {}, "annotations/train.json", "images/train")
register_coco_instances("nvr_val", {}, "annotations/val.json", "images/val")

TensorFlow/Keras

import tensorflow_datasets as tfds

# Load COCO format dataset
dataset = tfds.load('coco', data_dir='path/to/dataset')

Original Folder Structure Preserved

All original folder classification information is preserved in the annotation metadata:

• camera_location: Which camera captured the image
• primary_category: Main classification (sleeping_position, event, reference_subject)
• subcategory: Specific category within primary
• specific_event: Detailed event classification where applicable