OpenDetection-50K-Remastered

OpenDetection-50K-Remastered is an object detection dataset built primarily from general, publicly available images, which make up the major part of the input imagery, along with other publicly available datasets including the COCO dataset and Blip3o. General, all-class object detection annotations were produced using a custom-built RF-DETR computer vision pipeline driven by an object-manager task, yielding per-image object labels, confidence scores, and bounding boxes at scale.

Dataset Details

• Images: 52,638, primarily general publicly available images forming the major part of the dataset, along with other publicly available datasets including the COCO dataset and Blip3o
• Annotations: object labels, label IDs, confidence scores, and bounding boxes, produced via a custom RF-DETR detection pipeline
• Format: parquet, optimized-parquet
• Split: train (52,638 rows)
• Total size: 32.6 GB
• Language: English
• License: Apache 2.0

Source Data

The major part of the input imagery consists of general, publicly available images. This is supplemented by other publicly available datasets, including the COCO dataset and Blip3o, to build a broad and diverse pool of images for detection.

Annotation Pipeline

Objects were detected using a custom-built pipeline based on RF-DETR. An object-manager task orchestrates detection across all images to produce general, all-object annotations for each image, covering multiple object classes per image with associated confidence scores and bounding box coordinates.

Compute Infrastructure

Dataset generation was run using Hugging Face Jobs, which provided the compute infrastructure supporting large-scale annotation across the full image set.

Dataset Structure

The dataset contains a single train split with the following fields.

Field	Type	Description
image	image	Source input image
objects	string	JSON-encoded list of detections, each with label, label_id, score, and bbox
annotated_image	image	Source image rendered with bounding box annotations overlaid

Uses

This dataset can be used for training or fine-tuning object detection models, benchmarking detection pipelines, and building open-vocabulary or auto-annotation workflows that require labeled bounding boxes across diverse image sources.

Licensing

This dataset is released under the Apache 2.0 license. Users should also refer to the terms of use for the underlying source imagery and datasets.

Acknowledgements

• COCO dataset
• Blip3o
• RF-DETR
• Hugging Face Jobs for compute infrastructure supporting dataset generation, thanks for the compute support

Citation

@misc{prithiv_sakthi_2026,
    author       = { Prithiv Sakthi },
    title        = { OpenDetection-50K-Remastered (Revision ea85cab) },
    year         = 2026,
    url          = { https://huggingface.co/datasets/prithivMLmods/OpenDetection-50K-Remastered },
    doi          = { 10.57967/hf/9516 },
    publisher    = { Hugging Face }
}