Roboflow Adds Video Mirroring from S3, GCS and Azure

Roboflow has added video mirroring from AWS S3, S3-compatible storage, Google Cloud Storage and Azure Blob Storage to its Datasources feature. Teams can connect cloud buckets, sync selected MP4 and MOV files, extract representative frames and prepare them for annotation.
- Roboflow Datasources supports video mirroring from AWS S3, S3-compatible storage, Google Cloud Storage and Azure Blob Storage.
- Supported files are MP4 and QuickTime (.mov), with a maximum size of 3 GB per file.
- Daily scheduled syncing runs every 24 hours and is off by default.
- Users must extract and select frames separately before annotation and dataset creation.
- Roboflow says teams can review labels, create a dataset version and train an RF-DETR model.
Cloud footage can flow into a Roboflow workspace
Roboflow Datasources creates a saved connection between cloud storage and a Roboflow workspace. The connection combines a bucket or container, reusable credentials, file-selection rules and a sync configuration. With scheduling enabled, Roboflow checks for matching data automatically and stores mirrored videos in the Asset Library.
The feature supports MP4 and QuickTime (.mov) files from AWS S3, S3-compatible storage, Google Cloud Storage and Azure Blob Storage. The maximum size is 3 GB per file for each provider. Files larger than 3 GB must be split into smaller clips before syncing.
According to Roboflow, the capability is aimed at teams working with footage from multiple cameras, locations and operating conditions. Examples in the company’s announcement include manufacturing teams collecting footage from several production lines and logistics teams gathering video from multiple facilities.
Scheduled syncing avoids repeated manual imports
Users begin by saving access credentials in their workspace, then create a Datasource and provide the relevant bucket or container details. File filters based on glob patterns can narrow the import to particular folders, cameras, vehicles or production lines. For example, a team could select MP4 files in a line-3 folder or footage associated with one vehicle.
Roboflow recommends starting with a narrow selection, such as one camera, production line or route. This allows teams to inspect the first import and reduce unnecessary data transfer, storage and annotation work. Once the selection is confirmed, users can enable a daily schedule that runs every 24 hours. Manual syncing is also available.
The announcement describes separate access and review controls for production data. Teams can use scoped cloud permissions, role-based access controls, custom roles and folder permissions. Audit logs record significant workspace actions, including image approvals, dataset exports and training activity. Roboflow also advises preserving information such as site, camera, capture time and product type when datasets span several facilities.
Video mirroring is only the first dataset step
Mirroring a video into the Asset Library does not automatically create a training dataset. Roboflow treats frame extraction and annotation as separate steps because a video can contain many nearly identical images. The company notes that one minute of video at 30 frames per second contains 1,800 frames, making selective sampling important for annotation efficiency.
To create an object detection dataset, users select source clips, open an Object Detection Project and add the clips through the web uploader. Roboflow’s video extraction dialog allows users to set an interval, trim a range and select individual frames. The sampling rate should reflect how quickly useful information changes in the footage.
After selecting frames, teams review their labels, create a dataset version and can train an RF-DETR model for their application. This workflow gives users a way to bring newly arriving cloud footage into a workspace while choosing which scenes and conditions should expand the training data.


