project.md

:::roboflow.core.project

Upload a native Action Recognition video

Project.upload_video sends the original MP4 or MOV bytes to a signed upload URL. It creates a video Source in the project; it does not extract frames or run inference. The platform processes the upload asynchronously.

project = rf.workspace("my-workspace").project("my-actions")
status = project.upload_video(
    "clip.mp4",
    batch_name="session-1",
    tag_names=["indoor"],
    metadata={"camera": "front"},
    split="train",
)
if status["status"] == "pending":
    status = project.wait_for_video_upload(status["videoId"], poll_timeout=300)

if status["status"] == "failed":
    raise RuntimeError(status["message"])

source_id = status["videoId"]  # Use this Source ID for video annotations.

upload_video(..., wait=True) performs the bounded wait in one call. The returned status is the API response: pending, uploaded (with resolvedBatch), or failed (with message). Poll later with project.get_video_upload_status(video_id). Always use videoId from the final uploaded response because ingestion can deduplicate onto another Source. Batch, tags, metadata, and split follow the platform upload API; the API validates their values. A timeout leaves the upload running, so poll its original ID later. poll_timeout=0 makes one status request and returns a terminal result if available. Status requests use the remaining polling budget as their connection and read inactivity timeout; this is not a strict whole-response wall-clock limit for a slowly streaming server.