Using video footage with Atlas

Use video footage as a source of still images for Atlas. You can reconstruct a mostly static environment from selected frames, or generate a camera move around a frozen moment captured by synchronized cameras.

The customer beta takes images for these workflows. Direct video upload and time control across a changing scene are not currently available.

Reconstruct an environment from a walkthrough

Use this workflow when the camera moves through a room or another mostly static space.

1. Select a short, continuous part of the capture. Keep the scene, lighting, and exposure consistent. Choose footage where the camera moves between positions so you see surfaces from different angles.
2. Export selected frames as images. Use a video editor or frame-extraction tool to save sharp, overlapping views. For a single room, start with 8–16 well-distributed images. Add views where they show an unseen area or another side of an object; the count is a starting point, not an API limit.
3. Reconstruct the images together. In Viewfinder, import the selected images together with Reconstruct. Through the Developer API, send them in one images2PosedRGBD request. This estimates their relative camera positions and depth in one scene.
4. Inspect the recovered scene. Check that neighboring views line up and cover the area you want to explore. Replace blurred or inconsistent images and add coverage where needed before generating new views.

Choose frames by sharpness, overlap, and coverage. A fixed extraction interval can be a starting point, but inspect the selected images: camera speed changes how much adjacent frames overlap. Uploading every video frame can add many near-duplicates without revealing more of the scene.

Aim for roughly 50% shared scene content between neighboring views. Include the floor, walls, lower furniture, and areas behind obstructions. See Splat best practices for capture examples and coverage guidance.

Images taken at different times work here because the scene stays mostly static. Moving people or objects can appear in incompatible positions across views. Avoid using those frames to reconstruct one shared scene. These recipes use perspective images; convert 360-degree or fisheye footage to perspective views before using the image-based workflow.

Choose your output

For a rendered walkthrough, author a camera path in Viewfinder, generate, and export the video. A finished splat is not required for this step.

Through the API, pass the returned posed views to atlasGenerate as contextFrames, with the complete ordered camera path as targetCameras in one request. Assemble the returned images at the frame rate used to sample that path. Independent requests do not share generated sequence context and can produce flicker or layout changes; see Generate a video or camera path.

For a navigable 3D asset, continue with Posed images to splats. Inspect the reconstructed views and their coverage first. Atlas can invent uncaptured areas, and estimated geometry is not a guarantee of measured dimensions.

Make a frozen-time camera move

Use synchronized cameras with overlapping views of the same event. Each camera must show the scene at the same instant; successive frames from one camera moving around a changing subject do not provide those simultaneous views.

1. Choose the instant to freeze. Align the recordings in time, then export one image from each camera at that instant. Matching frame numbers alone are insufficient if the recordings started at different times.
2. Reconstruct those images together. Submit the synchronized images in one images2PosedRGBD request and inspect the recovered views.
3. Generate the virtual camera move. Pass those posed views to atlasGenerate with your full camera path. Start and end at the physical camera viewpoints you want to join to the original footage.
4. Edit the result into your video. Play the original footage up to the chosen instant, insert the generated camera move, then resume the original footage from the ending camera at that instant.

This creates a camera move around one frozen scene. You choose the freeze point and assemble the surrounding footage in a video editor. The beta does not expose continuous reframing while the captured scene changes over time.

Test a short move first and inspect the subject from the intended viewpoints. Human faces, articulated subjects, and fine details may change in generated views. See Best practices for current quality limitations.