Frame

Attempts to acquire an image from the camera that corresponds to the current frame. Depending on device performance, can throw NotYetAvailableException for several frames after session start, and for a few frames at a time while the session is running.

Attempts to acquire a depth Android Image object that corresponds to the current frame.

The depth image has a single 16-bit plane at index 0, stored in little-endian format. Each pixel contains the distance in millimeters to the camera plane. Currently, the three most significant bits are always set to 000. The remaining thirteen bits express values from 0 to 8191, representing depth in millimeters. To extract distance from a depth map, see the Depth API developer guide .

The actual size of the depth image depends on the device and its display aspect ratio. The size of the depth image is typically around 160x120 pixels, with higher resolutions up to 640x480 on some devices. These sizes may change in the future. The outputs of acquireDepthImage() , acquireRawDepthImage() and acquireRawDepthConfidenceImage() will all have the exact same size.

Optimal depth accuracy occurs between 500 millimeters (50 centimeters) and 5000 millimeters (5 meters) from the camera. Error increases quadratically as distance from the camera increases.

Depth is estimated using data from the world-facing cameras, user motion, and hardware depth sensors such as a time-of-flight sensor (or ToF sensor) if available. As the user moves their device through the environment, 3D depth data is collected and cached which improves the quality of subsequent depth images and reducing the error introduced by camera distance.

If an up-to-date depth image isn't ready for the current frame, the most recent depth image available from an earlier frame will be returned instead. This is expected only to occur on compute-constrained devices. An up-to-date depth image should typically become available again within a few frames.

The image must be released via Image.close() once it is no longer needed.

Attempts to acquire a depth Android Image object that corresponds to the current frame.

The depth image has format HardwareBuffer.D_16 , which is a single 16-bit plane at index 0, stored in little-endian format. Each pixel contains the distance in millimeters to the camera plane, with the representable depth range between 0 millimeters and 65535 millimeters, or about 65 meters.

To extract distance from a depth map, see the Depth API developer guide .

The actual size of the depth image depends on the device and its display aspect ratio. The size of the depth image is typically around 160x120 pixels, with higher resolutions up to 640x480 on some devices. These sizes may change in the future. The outputs of acquireDepthImage16Bits() , acquireRawDepthImage16Bits() and acquireRawDepthConfidenceImage() will all have the exact same size.

Optimal depth accuracy occurs between 500 millimeters (50 centimeters) and 15000 millimeters (15 meters) from the camera, with depth reliably observed up to 25000 millimeters (25 meters). Error increases quadratically as distance from the camera increases.

Depth is estimated using data from the world-facing cameras, user motion, and hardware depth sensors such as a time-of-flight sensor (or ToF sensor) if available. As the user moves their device through the environment, 3D depth data is collected and cached which improves the quality of subsequent depth images and reducing the error introduced by camera distance. The depth accuracy improves as the user moves more while looking at a desired location.

If an up-to-date depth image isn't ready for the current frame, the most recent depth image available from an earlier frame will be returned instead. This is expected only to occur on compute-constrained devices. An up-to-date depth image should typically become available again within a few frames.

When the Geospatial API and the Depth API are enabled, output images from the Depth API will include terrain and building geometry when in a location with VPS coverage. See the Geospatial Depth Developer Guide for more information.

The image must be released via Image.close() once it is no longer needed.

Acquires the current set of estimated 3d points attached to real-world geometry. PointCloud.release() must be called after application is done using the PointCloud object.

Note: This information is for visualization and debugging purposes only. Its characteristics and format are subject to change in subsequent versions of the API.

Attempts to acquire the confidence Android Image object corresponding to the raw depth image of the current frame.

The image must be released via Image.close() once it is no longer needed.

Each pixel is an 8-bit unsigned integer representing the estimated confidence of the corresponding pixel in the raw depth image. The confidence value is between 0 and 255, inclusive, with 0 representing the lowest confidence and 255 representing the highest confidence in the measured depth value. Pixels without a valid depth estimate have a confidence value of 0 and a corresponding depth value of 0 (see acquireRawDepthImage() ).

The scaling of confidence values is linear and continuous within this range. Expect to see confidence values represented across the full range of 0 to 255, with values increasing as better observations are made of each location. If an application requires filtering out low-confidence pixels, removing depth pixels below a confidence threshold of half confidence (128) tends to work well.

The actual size of the depth image depends on the device and its display aspect ratio. The size of the depth image is typically around 160x120 pixels, with higher resolutions up to 640x480 on some devices. These sizes may change in the future. The outputs of acquireDepthImage() , acquireRawDepthImage() and acquireRawDepthConfidenceImage() will all have the exact same size.

Attempts to acquire a "raw", mostly unfiltered, depth Android Image object that corresponds to the current frame.

The raw depth image is sparse and does not provide valid depth for all pixels. Pixels without a valid depth estimate have a pixel value of 0 and a corresponding confidence value of 0 (see acquireRawDepthConfidenceImage() ).

The depth image has a single 16-bit plane at index 0, stored in little-endian format. Each pixel contains the distance in millimeters to the camera plane. Currently, the three most significant bits are always set to 000. The remaining thirteen bits express values from 0 to 8191, representing depth in millimeters. To extract distance from a depth map, see the Depth API developer guide .

The actual size of the depth image depends on the device and its display aspect ratio. The size of the depth image is typically around 160x120 pixels, with higher resolutions up to 640x480 on some devices. These sizes may change in the future. The outputs of acquireDepthImage() , acquireRawDepthImage() and acquireRawDepthConfidenceImage() will all have the exact same size.

Optimal depth accuracy occurs between 500 millimeters (50 centimeters) and 5000 millimeters (5 meters) from the camera. Error increases quadratically as distance from the camera increases.

Depth is primarily estimated using data from the motion of world-facing cameras. As the user moves their device through the environment, 3D depth data is collected and cached, improving the quality of subsequent depth images and reducing the error introduced by camera distance. Depth accuracy and robustness improves if the device has a hardware depth sensor, such as a time-of-flight (ToF) camera.

Not every raw depth image contains a new depth estimate. Typically there are about 10 updates to the raw depth data per second. The depth images between those updates are a 3D reprojection which transforms each depth pixel into a 3D point in space and renders those 3D points into a new raw depth image based on the current camera pose. This effectively transforms raw depth image data from a previous frame to account for device movement since the depth data was calculated. For some applications it may be important to know whether the raw depth image contains new depth data or is a 3D reprojection (for example, to reduce the runtime cost of 3D reconstruction). To do that, compare the current raw depth image timestamp, obtained via Image.getTimestamp() , with the previously recorded raw depth image timestamp. If they are different, the depth image contains new information.

The image must be released via Image.close() once it is no longer needed.

Attempts to acquire a "raw", mostly unfiltered, depth Android Image object that corresponds to the current frame.

The raw depth image is sparse and does not provide valid depth for all pixels. Pixels without a valid depth estimate have a pixel value of 0 and a corresponding confidence value of 0 (see acquireRawDepthConfidenceImage() ).

The depth image has format HardwareBuffer.D_16 , which is a single 16-bit plane at index 0, stored in little-endian format. Each pixel contains the distance in millimeters to the camera plane, with the representable depth range between 0 millimeters and 65535 millimeters, or about 65 meters.

To extract distance from a depth map, see the Depth API developer guide .

The actual size of the depth image depends on the device and its display aspect ratio. The size of the depth image is typically around 160x120 pixels, with higher resolutions up to 640x480 on some devices. These sizes may change in the future. The outputs of acquireDepthImage16Bits() , acquireRawDepthImage16Bits() and acquireRawDepthConfidenceImage() will all have the exact same size.

Optimal depth accuracy occurs between 500 millimeters (50 centimeters) and 15000 millimeters (15 meters) from the camera, with depth reliably observed up to 25000 millimeters (25 meters). Error increases quadratically as distance from the camera increases.

Depth is primarily estimated using data from the motion of world-facing cameras. As the user moves their device through the environment, 3D depth data is collected and cached, improving the quality of subsequent depth images and reducing the error introduced by camera distance. The depth accuracy improves as the user moves more while looking at a desired location. Depth accuracy and robustness improves if the device has a hardware depth sensor, such as a time-of-flight (ToF) camera.

Not every raw depth image contains a new depth estimate. Typically there are about 10 updates to the raw depth data per second. The depth images between those updates are a 3D reprojection which transforms each depth pixel into a 3D point in space and renders those 3D points into a new raw depth image based on the current camera pose. This effectively transforms raw depth image data from a previous frame to account for device movement since the depth data was calculated. For some applications it may be important to know whether the raw depth image contains new depth data or is a 3D reprojection (for example, to reduce the runtime cost of 3D reconstruction). To do that, compare the current raw depth image timestamp, obtained via Image.getTimestamp() , with the previously recorded raw depth image timestamp. If they are different, the depth image contains new information.

When the Geospatial API and the Depth API are enabled, output images from the Depth API will include terrain and building geometry when in a location with VPS coverage. See the Geospatial Depth Developer Guide for more information.

The image must be released via Image.close() once it is no longer needed.

Attempts to acquire the semantic confidence image corresponding to the current frame. Each pixel is an 8-bit integer representing the estimated confidence of the corresponding pixel in the semantic image. See the Scene Semantics Developer Guide for more information.

The confidence value is between 0 and 255, inclusive, with 0 representing the lowest confidence and 255 representing the highest confidence in the semantic class prediction from acquireSemanticImage() .

The image must be released via Image.close() once it is no longer needed.

In order to obtain a valid result from this function, you must set the session's Config.SemanticMode to Config.SemanticMode.ENABLED . Use Session.isSemanticModeSupported(Config.SemanticMode) to query for support for Scene Semantics.

The size of the semantic confidence image is the same size as the image obtained by acquireSemanticImage() .

Attempts to acquire the semantic image corresponding to the current frame. Each pixel in the image is an 8-bit unsigned integer representing a semantic class label: see SemanticLabel for a list of pixel labels and the Scene Semantics Developer Guide for more information.

The image must be released via Image.close() once it is no longer needed.

In order to obtain a valid result from this function, you must set the session's Config.SemanticMode to Config.SemanticMode.ENABLED . Use Session.isSemanticModeSupported(Config.SemanticMode) to query for support for Scene Semantics.

The width of the semantic image is currently 256 pixels. The height of the image depends on the device and will match its display aspect ratio.

Returns the ( Android Camera timestamp ) of the image.

Returns the pose of the Android Sensor Coordinate System in the world coordinate space for this frame. The orientation follows the device's "native" orientation (it is not affected by display rotation) with all axes corresponding to those of the Android sensor coordinates.

See Also:

• Camera.getPose() for the pose of the physical camera.
• Camera.getDisplayOrientedPose() for the pose of the virtual camera.

Returns the Camera object for the session. Note that Camera instances are long-lived and may be kept across frames for the duration of the session. Repeated calls to this method will return distinct Camera objects that refer to the same underlying camera state and are equal to each other.

Returns the OpenGL ES camera texture name (id) associated with this frame. This is guaranteed to be one of the texture names previously set via Session.setCameraTextureNames(int[]) or Session.setCameraTextureName(int) . Texture names (ids) are returned in a round robin fashion in sequential frames.

Gets the HardwareBuffer for this frame. See Vulkan Rendering developer guide for more information.

Should only be called when a configuration is active that uses Config.TextureUpdateMode.EXPOSE_HARDWARE_BUFFER .

Returns the camera metadata for the current camera image, if available. Throws NotYetAvailableException when metadata is not yet available due to sensors data not yet being available.

If the AR session was created for shared camera access, this method will throw IllegalStateException . To retrieve image metadata in shared camera mode, use SharedCamera.setCaptureCallback(CameraCaptureSession.CaptureCallback, Handler) , then use getAndroidCameraTimestamp() to correlate the frame to metadata retrieved from CameraCaptureSession.CaptureCallback .

Returns the current ambient light estimate, if light estimation was enabled.

If lighting estimation is not enabled in the session configuration, the returned LightingEstimate will always return LightEstimate.State.NOT_VALID from LightEstimate.getState() .

Retrieves the percentage of pixels in the most recent semantics image that are queryLabel .

Queries the semantic image provided by acquireSemanticImage() for pixels labelled by queryLabel . This call is more efficient than retrieving the Image and performing a pixel-wise search for the detected label.

Returns the timestamp in nanoseconds when this image was captured. This can be used to detect dropped frames or measure the camera frame rate. The time base of this value is specifically not defined, but it is likely similar to System.nanoTime() .

Returns the anchors that were changed by the Session.update() that returned this Frame.

Retrieve all track data that was written to the specified track during the current frame. If frames are skipped during playback, which can happen when the device is under load, played back track data will be attached to a later frame in order.

Each call to recordTrackData(UUID, ByteBuffer) at recording time will be returned as a separate TrackData entry in the collection.

Returns the trackables of a particular type that were changed by the Session.update() that returned this Frame. filterType may be Plane.class or Point.class , or Trackable.class to retrieve all changed trackables.

Checks if the display rotation or viewport geometry changed since the previous Frame . The application should re-query Camera.getProjectionMatrix(float[], int, float, float) and transformCoordinates2d(Coordinates2d, float[], Coordinates2d, float[]) whenever this is true.

Similar to hitTest(float, float) , but will take values from Android MotionEvent . It is assumed that the MotionEvent is received from the same view that was used as the size for Session.setDisplayGeometry(int, int, int) .

Note: this method does not consider the action of the MotionEvent . The caller must check for appropriate action, if needed, before calling this method.

Note: When using Session.Feature.FRONT_CAMERA , the returned hit result list will always be empty, as the camera is not TrackingState.TRACKING . Hit testing against tracked faces is not currently supported.

Similar to hitTest(float, float) , but takes an arbitrary ray in world space coordinates instead of a screen-space point.

Note: When using Session.Feature.FRONT_CAMERA , the returned hit result list will always be empty, as the camera is not TrackingState.TRACKING . Hit testing against tracked faces is not currently supported.

Performs a ray cast from the user's device in the direction of the given location in the camera view. Intersections with detected scene geometry are returned, sorted by distance from the device; the nearest intersection is returned first.

Note: Significant geometric leeway is given when returning hit results. For example, a plane hit may be generated if the ray came close, but did not actually hit within the plane extents or plane bounds ( Plane.isPoseInExtents(Pose) and Plane.isPoseInPolygon(Pose) can be used to determine these cases). A point (point cloud) hit is generated when a point is roughly within one finger-width of the provided screen coordinates.

Note: When using Session.Feature.FRONT_CAMERA , the returned hit result list will always be empty, as the camera is not TrackingState.TRACKING . Hit testing against tracked faces is not currently supported.

Note: In ARCore 1.24.0 or later on supported devices, if depth is enabled by calling Config.setDepthMode(Config.DepthMode) with the value Config.DepthMode.AUTOMATIC , the returned list includes DepthPoint values sampled from the latest computed depth image.

Performs a ray cast that can return a result before ARCore establishes full tracking.

The pose and apparent scale of attached objects depends on the InstantPlacementPoint tracking method and the provided approximateDistanceMeters. A discussion of the different tracking methods and the effects of apparent object scale are described in InstantPlacementPoint .

This function will succeed only if Config.InstantPlacementMode is Config.InstantPlacementMode.LOCAL_Y_UP in the ARCore session configuration, the ARCore session tracking state is TrackingState.TRACKING , and there are sufficient feature points to track the point in screen space.

Writes a data sample in the specified track. The samples recorded using this API are muxed into the recorded MP4 dataset as an additional MP4 stream.

Multiple samples can be recorded to the same frame and will be played back together.

For smooth playback of the MP4 on video players and for future compatibility of the MP4 datasets with ARCore's playback of data tracks it is recommended that the samples are recorded at a frequency no higher than 90kHz.

Additionally, if the samples are recorded at a frequency lower than 1Hz, empty (zero byte) padding samples will be automatically recorded at approximately one second intervals to fill in the gaps.

Recording samples introduces additional CPU and/or I/O overhead and may affect app performance.

Transforms a list of 2D coordinates from one 2D coordinate system to another 2D coordinate system.

For Android view coordinates ( Coordinates2d.VIEW , Coordinates2d.VIEW_NORMALIZED ), the view information is taken from the most recent call to Session.setDisplayGeometry(int, int, int) .

Must be called on the most recently obtained Frame object. If this function is called on an older frame, a log message will be printed and outputVertices2d will remain unchanged.

Some examples of useful conversions:

• To transform from [0,1] range to screen-quad coordinates for rendering: Coordinates2d.VIEW_NORMALIZED -> Coordinates2d.TEXTURE_NORMALIZED
• To transform from [-1,1] range to screen-quad coordinates for rendering: Coordinates2d.OPENGL_NORMALIZED_DEVICE_COORDINATES -> Coordinates2d.TEXTURE_NORMALIZED
• To transform a point found by a computer vision algorithm in a CPU image into a point on the screen that can be used to place an Android View (e.g. Button) at that location: Coordinates2d.IMAGE_PIXELS -> Coordinates2d.VIEW
• To transform a point found by a computer vision algorithm in a CPU image into a point to be rendered using GL in clip-space ( [-1,1] range): Coordinates2d.IMAGE_PIXELS -> Coordinates2d.OPENGL_NORMALIZED_DEVICE_COORDINATES

Read-only array-backed buffers are not supported by inputVertices2d for performance reasons.

If inputCoordinates is same as outputCoordinates , the input vertices will be copied to the output vertices unmodified.

Transforms a list of 2D coordinates from one 2D coordinate system to another 2D coordinate system.

Same as transformCoordinates2d(Coordinates2d, FloatBuffer, Coordinates2d, FloatBuffer) , but taking float arrays.

Transforms a list of 2D coordinates from one 2D coordinate system to 3D coordinate space. See the Electronic Image Stabilization developer guide for more information.

The view information is taken from the most recent call to Session.setDisplayGeometry(int, int, int) .

If Electronic Image Stabilization is off, the device coordinates return (-1, -1, 0) -> (1, 1, 0) and texture coordinates return the same coordinates as transformCoordinates2d(Coordinates2d, float[], Coordinates2d, float[]) } with the Z component set to 1.0f.

In order to use EIS, your app should use EIS compensated screen coordinates and camera texture coordinates to pass on to shaders. Use the 2D NDC space coordinates as input to obtain EIS compensated 3D screen coordinates and matching camera texture coordinates.

Transforms a list of 2D coordinates from one 2D coordinate system to another 3D coordinate system.

Same as transformCoordinates3d(Coordinates2d, FloatBuffer, Coordinates3d, FloatBuffer) , but taking float arrays.

Transform the given texture coordinates to correctly show the background image. This will account for the display rotation, and any additional required adjustment. For performance, this function should be called only if hasDisplayGeometryChanged() returns true.

Usage Notes / Bugs:

• Both input and output buffers must be direct and native byte order.
• Position and limit of buffers is ignored.
• Capacity of both buffers must be identical.
• Capacity of both buffers must be a multiple of 2.

Note: both buffer positions will remain unmodified after this call.