HOT3DHUGGFACE/data_loaders/AriaDataProvider.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import os
from typing import Dict, List, Optional

import numpy as np
from data_loaders.frameset import compute_frameset_for_timestamp
from projectaria_tools.core import data_provider  # @manual
from projectaria_tools.core.calibration import (  # @manual
    CameraCalibration,
    DeviceCalibration,
    distort_by_calibration,
    FISHEYE624,
    get_linear_camera_calibration,
    LINEAR,
)
from projectaria_tools.core.mps import (  # @manual
    EyeGaze,
    get_eyegaze_point_at_depth,
    MpsDataPathsProvider,
    MpsDataProvider,
)
from projectaria_tools.core.sensor_data import TimeDomain, TimeQueryOptions  # @manual
from projectaria_tools.core.sophus import SE3  # @manual
from projectaria_tools.core.stream_id import StreamId  # @manual


class AriaDataProvider:
    def __init__(
        self, vrs_filepath: str, mps_folder_path: Optional[str] = None
    ) -> None:
        self._vrs_data_provider = data_provider.create_vrs_data_provider(vrs_filepath)

        # MPS data provider
        if mps_folder_path is not None and os.path.exists(mps_folder_path):
            mps_data_paths_provider = MpsDataPathsProvider(mps_folder_path)
            mps_data_paths = mps_data_paths_provider.get_data_paths()
            self._mps_data_provider = MpsDataProvider(mps_data_paths)
            print(mps_data_paths)
        else:
            self._mps_data_provider = None

        # Pre-compute the sorted timestamps for each stream
        self._stream_timestamps_sorted: Dict[str, List[int]] = {}
        for stream_id in self.get_image_stream_ids():
            self._stream_timestamps_sorted[str(stream_id)] = sorted(
                self.get_sequence_timestamps(stream_id, TimeDomain.TIME_CODE)
            )

    def get_image_stream_ids(self) -> List[StreamId]:
        # retrieve all streams ids and filter the one that are image based
        stream_ids = self._vrs_data_provider.get_all_streams()
        image_stream_ids = [
            p
            for p in stream_ids
            if self._vrs_data_provider.get_label_from_stream_id(p).startswith("camera-")
        ]
        return image_stream_ids

    def get_sequence_timestamps(
        self,
        stream_id: StreamId = StreamId("214-1"),
        time_domain: TimeDomain = TimeDomain.TIME_CODE,
    ) -> List[int]:
        """
        Returns the list of "time code" timestamp for the sequence
        """
        return self._vrs_data_provider.get_timestamps_ns(stream_id, time_domain)

    def get_frameset_from_timestamp(
        self,
        timestamp_ns: int,
        frameset_acceptable_time_diff_ns: int,
        time_domain: TimeDomain = TimeDomain.TIME_CODE,
    ) -> Dict[str, Optional[int]]:
        """
        Computes a frameset from a given timestamp within an acceptable time difference.
        The frameset consists of the closest timestamps for each stream that are within the acceptable time difference.
        For Aria, the recommended acceptable time difference is 1e6 ns (or 1ms).
        Returns a dictionary mapping each str(StreamId) to its closest timestamp.
        """
        if time_domain is not TimeDomain.TIME_CODE:
            raise ValueError(
                f"{time_domain} is not supported. Only TIME_CODE is supported"
            )
        out_frameset = compute_frameset_for_timestamp(
            stream_timestamps_sorted=self._stream_timestamps_sorted,
            target_timestamp=timestamp_ns,
            frameset_acceptable_time_diff=frameset_acceptable_time_diff_ns,
        )
        return out_frameset

    def get_image_stream_label(self, stream_id: StreamId) -> str:
        return self._vrs_data_provider.get_label_from_stream_id(stream_id)

    def get_image(self, timestamp_ns: int, stream_id: StreamId) -> np.ndarray:
        image = self._vrs_data_provider.get_image_data_by_time_ns(
            stream_id,
            timestamp_ns,
            TimeDomain.TIME_CODE,
            TimeQueryOptions.CLOSEST,
        )
        return image[0].to_numpy_array() if image is not None else None

    def get_undistorted_image(
        self, timestamp_ns: int, stream_id: StreamId
    ) -> np.ndarray:
        image = self.get_image(timestamp_ns, stream_id)

        [T_device_camera, native_camera_online_calibration] = (
            self.get_online_camera_calibration(
                stream_id, timestamp_ns=timestamp_ns, camera_model=FISHEYE624
            )
        )
        [T_device_camera, pinhole_camera_online_calibration] = (
            self.get_online_camera_calibration(
                stream_id, timestamp_ns=timestamp_ns, camera_model=LINEAR
            )
        )

        # Compute the actual undistorted image
        undistorted_image = distort_by_calibration(
            image, pinhole_camera_online_calibration, native_camera_online_calibration
        )

        return undistorted_image

    def get_device_calibration(self) -> DeviceCalibration:
        """
        Return the device calibration (factory calibration of all sensors)
        """
        return self._vrs_data_provider.get_device_calibration()

    def get_camera_calibration(
        self,
        stream_id: StreamId,
        camera_model=FISHEYE624,
    ) -> tuple[SE3, CameraCalibration]:
        """
        Return the camera calibration of the device of the sequence as [Extrinsics, Intrinsics]
        Note:
         - A corresponding pinhole camera can be requested by using camera_model = LINEAR.
         - This is the camera model used to generate the 'get_undistorted_image'.
        """
        if not (camera_model is FISHEYE624 or camera_model is LINEAR):
            raise ValueError(
                "Invalid camera_model type, only FISHEYE624 and LINEAR are supported"
            )

        device_calibration = self.get_device_calibration()
        stream_label = self._vrs_data_provider.get_label_from_stream_id(stream_id)
        camera_calibration = device_calibration.get_camera_calib(stream_label)

        # Store the relative transform from device to camera
        T_device_camera = camera_calibration.get_transform_device_camera()

        # If a corresponding pinhole camera is requested, we build one on the fly
        if camera_model == LINEAR:
            focal_lengths = camera_calibration.get_focal_lengths()
            image_size = camera_calibration.get_image_size()
            camera_calibration = get_linear_camera_calibration(
                image_size[0], image_size[1], focal_lengths[0]
            )
        # else return the native FISHEYE624 camera model

        return [T_device_camera, camera_calibration]

    def get_online_camera_calibration(
        self,
        stream_id: StreamId,
        timestamp_ns: Optional[int],
        time_domain: TimeDomain = TimeDomain.TIME_CODE,
        camera_model=FISHEYE624,
    ) -> tuple[SE3, CameraCalibration]:
        """
        Return the camera calibration of the device of the sequence as [Extrinsics, Intrinsics]
        Note:
         - A corresponding pinhole camera can be requested by using camera_model = LINEAR.
         - This is the camera model used to generate the 'get_undistorted_image'.
        """
        if not (camera_model is FISHEYE624 or camera_model is LINEAR):
            raise ValueError(
                "Invalid camera_model type, only FISHEYE624 and LINEAR are supported"
            )
        if time_domain is not TimeDomain.TIME_CODE:
            raise ValueError(
                f"{time_domain} is not supported. Only TIME_CODE is supported"
            )

        device_timestamp_ns = (
            self._vrs_data_provider.convert_from_timecode_to_device_time_ns(
                timestamp_ns
            )
        )
        online_calibration = self._mps_data_provider.get_online_calibration(
            device_timestamp_ns=device_timestamp_ns,
            time_query_options=TimeQueryOptions.CLOSEST,
        )
        camera_calibs = online_calibration.camera_calibs

        stream_label = self._vrs_data_provider.get_label_from_stream_id(stream_id)
        camera_calib = [c for c in camera_calibs if c.get_label() == stream_label]
        if len(camera_calib) == 0:
            raise ValueError(
                f"camera_calib not found for stream_label: {stream_label} stream_id: {stream_id} at timestamp_ns: {timestamp_ns}"
            )
        camera_calibration = camera_calib[0]

        ## Fix the image size to correspond to the image saved in the vrs.
        ## The calibration returned by mps_data_provider has hardcoded image sizes which are incorrect.
        [_, native_camera_calibration] = self.get_camera_calibration(
            stream_id, camera_model=camera_model
        )
        camera_calibration = CameraCalibration(
            camera_calibration.get_label(),
            camera_model,
            camera_calibration.projection_params(),
            camera_calibration.get_transform_device_camera(),
            native_camera_calibration.get_image_size()[0],  ## correct the image size
            native_camera_calibration.get_image_size()[1],
            camera_calibration.get_valid_radius(),
            camera_calibration.get_max_solid_angle(),
            camera_calibration.get_serial_number(),
        )
        # Store the relative transform from device to camera
        T_device_camera = camera_calibration.get_transform_device_camera()

        # If a corresponding pinhole camera is requested, we build one on the fly
        if camera_model == LINEAR:
            focal_lengths = camera_calibration.get_focal_lengths()
            image_size = camera_calibration.get_image_size()
            camera_calibration = get_linear_camera_calibration(
                image_size[0], image_size[1], focal_lengths[0]
            )
            # Info: transform_device_camera is set to ID in this path in the camera_calibration
        # else return the native FISHEYE624 camera model

        return [T_device_camera, camera_calibration]

    def _timestamp_convert(
        self, timestamp: int, time_domain_in: TimeDomain, time_domain_out: TimeDomain
    ) -> int:
        """
        Returns the converted timestamp between two domains (TimeCode <-> Aria DeviceTime)
        """
        if self._vrs_data_provider:
            # Map to corresponding timestamp
            if (
                time_domain_in == TimeDomain.TIME_CODE
                and time_domain_out == TimeDomain.DEVICE_TIME
            ):
                out_timestamp = (
                    self._vrs_data_provider.convert_from_timecode_to_device_time_ns(
                        timestamp
                    )
                )
            if (
                time_domain_in == TimeDomain.DEVICE_TIME
                and time_domain_out == TimeDomain.TIME_CODE
            ):
                out_timestamp = (
                    self._vrs_data_provider.convert_from_device_time_to_timecode_ns(
                        timestamp
                    )
                )
            return out_timestamp
        return None

    ###
    # Add MPS data specifics
    ###

    def get_point_cloud(self) -> Optional[np.ndarray]:
        """
        Return the point cloud of the scene
        """
        if self._mps_data_provider is None:
            return None
        if self._mps_data_provider.has_semidense_point_cloud():
            point_cloud_data = self._mps_data_provider.get_semidense_point_cloud()
            # Point cloud filtering is left to the user
            return point_cloud_data

        return None

    def _get_gaze_vector_reprojection(
        self,
        eye_gaze: EyeGaze,
        stream_id_label: str,
        device_calibration: DeviceCalibration,
        camera_calibration: CameraCalibration,
    ) -> np.ndarray:
        """
        Helper function to project a eye gaze output onto a given image and its calibration, assuming specified fixed depth
        """
        gaze_center_in_cpf = get_eyegaze_point_at_depth(
            eye_gaze.yaw, eye_gaze.pitch, depth_m=eye_gaze.depth or 1.0
        )
        transform_device_cpf = device_calibration.get_transform_device_cpf()
        transform_device_camera = device_calibration.get_transform_device_sensor(
            stream_id_label, True
        )
        # We use CAD value (this is the coordinate system used by the Eye Gaze model prediction)
        # Using factory calibration (i.e CAD = False) would lead to less accurate EyeGaze reprojection.
        transform_camera_cpf = transform_device_camera.inverse() @ transform_device_cpf
        gaze_center_in_camera = transform_camera_cpf @ gaze_center_in_cpf
        gaze_center_in_pixels = camera_calibration.project(gaze_center_in_camera)
        return gaze_center_in_pixels

    def get_eye_gaze_in_camera(
        self,
        stream_id: StreamId,
        timestamp_ns: int,
        time_domain: TimeDomain = TimeDomain.TIME_CODE,
        camera_model=FISHEYE624,
    ):
        """
        Return the eye_gaze at the given timestamp projected in the given stream for the given time_domain
        """
        if not (camera_model is FISHEYE624 or camera_model is LINEAR):
            raise ValueError(
                "Invalid camera_model type, only FISHEYE624 and LINEAR are supported"
            )

        eye_gaze = self.get_eye_gaze(timestamp_ns, time_domain)
        if eye_gaze:
            [T_device_camera, camera_calibration] = self.get_camera_calibration(
                stream_id, camera_model
            )
            # Compute eye_gaze vector at depth_m and project it in the image
            gaze_projection = self._get_gaze_vector_reprojection(
                eye_gaze,
                self.get_image_stream_label(stream_id),
                self.get_device_calibration(),
                camera_calibration,
            )
            return gaze_projection
        return None

    def get_eye_gaze(
        self,
        timestamp_ns: int,
        time_domain: TimeDomain = TimeDomain.TIME_CODE,
    ) -> Optional[EyeGaze]:
        """
        Return the eye_gaze data at the given timestamp
        """
        # Map to corresponding timestamp
        if time_domain == TimeDomain.TIME_CODE:
            device_timestamp_ns = self._timestamp_convert(
                timestamp_ns, TimeDomain.TIME_CODE, TimeDomain.DEVICE_TIME
            )
        elif time_domain == TimeDomain.DEVICE_TIME:
            device_timestamp_ns = timestamp_ns
        else:
            raise ValueError("Unsupported time domain")

        if device_timestamp_ns:
            if self._mps_data_provider.has_personalized_eyegaze():
                return self._mps_data_provider.get_personalized_eyegaze(
                    device_timestamp_ns
                )
            elif self._mps_data_provider.has_general_eyegaze():
                return self._mps_data_provider.get_general_eyegaze(device_timestamp_ns)
        return None