mediapipe-rs/mediapipe/modules/hand_landmark/hand_landmark_tracking_cpu_image.pbtxt

# MediaPipe graph to detect/predict hand landmarks on CPU.
#
# The procedure is done in two steps:
# - locate palms/hands
# - detect landmarks for each palm/hand.
# This graph tries to skip palm detection as much as possible by reusing
# previously detected/predicted landmarks for new images.

type: "HandLandmarkTrackingCpuImage"

# Input image. (Image)
input_stream: "IMAGE:image"

# Max number of hands to detect/track. (int)
input_side_packet: "NUM_HANDS:num_hands"

# Complexity of hand landmark and palm detection models: 0 or 1. Accuracy as
# well as inference latency generally go up with the model complexity. If
# unspecified, functions as set to 1. (int)
input_side_packet: "MODEL_COMPLEXITY:model_complexity"

# Whether landmarks on the previous image should be used to help localize
# landmarks on the current image. (bool)
input_side_packet: "USE_PREV_LANDMARKS:use_prev_landmarks"

# The throttled input image. (Image)
output_stream: "IMAGE:throttled_image"

# Collection of detected/predicted hands, each represented as a list of
# landmarks. (std::vector<NormalizedLandmarkList>)
# NOTE: there will not be an output packet in the LANDMARKS stream for this
# particular timestamp if none of hands detected. However, the MediaPipe
# framework will internally inform the downstream calculators of the absence of
# this packet so that they don't wait for it unnecessarily.
output_stream: "LANDMARKS:multi_hand_landmarks"

# Collection of detected/predicted hand world landmarks.
# (std::vector<LandmarkList>)
#
# World landmarks are real-world 3D coordinates in meters with the origin in the
# center of the hand bounding box calculated from the landmarks.
#
# WORLD_LANDMARKS shares the same landmark topology as LANDMARKS. However,
# LANDMARKS provides coordinates (in pixels) of a 3D object projected onto the
# 2D image surface, while WORLD_LANDMARKS provides coordinates (in meters) of
# the 3D object itself.
output_stream: "WORLD_LANDMARKS:multi_hand_world_landmarks"

# Collection of handedness of the detected hands (i.e. is hand left or right),
# each represented as a ClassificationList proto with a single Classification
# entry. (std::vector<ClassificationList>)
# Note that handedness is determined assuming the input image is mirrored,
# i.e., taken with a front-facing/selfie camera with images flipped
# horizontally.
output_stream: "HANDEDNESS:multi_handedness"

# Extra outputs (for debugging, for instance).
# Detected palms. (std::vector<Detection>)
output_stream: "PALM_DETECTIONS:palm_detections"
# Regions of interest calculated based on landmarks.
# (std::vector<NormalizedRect>)
output_stream: "HAND_ROIS_FROM_LANDMARKS:hand_rects"
# Regions of interest calculated based on palm detections.
# (std::vector<NormalizedRect>)
output_stream: "HAND_ROIS_FROM_PALM_DETECTIONS:hand_rects_from_palm_detections"

node {
  calculator: "FlowLimiterCalculator"
  input_stream: "image"
  input_stream: "FINISHED:multi_hand_landmarks"
  input_stream_info: {
    tag_index: "FINISHED"
    back_edge: true
  }
  output_stream: "throttled_image"
  options: {
    [mediapipe.FlowLimiterCalculatorOptions.ext] {
      max_in_flight: 1
      max_in_queue: 1
    }
  }
}

# Converts Image to ImageFrame for HandLandmarkTrackingCpu to consume.
node {
  calculator: "FromImageCalculator"
  input_stream: "IMAGE:throttled_image"
  output_stream: "IMAGE_CPU:raw_image_frame"
  output_stream: "SOURCE_ON_GPU:is_gpu_image"
}

# TODO: Remove the extra flipping once adopting MlImage.
# If the source images are on gpu, flip the data vertically before sending them
# into HandLandmarkTrackingCpu. This maybe needed because OpenGL represents
# images assuming the image origin is at the bottom-left corner, whereas
# MediaPipe in general assumes the image origin is at the top-left corner.
node: {
  calculator: "ImageTransformationCalculator"
  input_stream: "IMAGE:raw_image_frame"
  input_stream: "FLIP_VERTICALLY:is_gpu_image"
  output_stream: "IMAGE:image_frame"
}

node {
  calculator: "HandLandmarkTrackingCpu"
  input_stream: "IMAGE:image_frame"
  input_side_packet: "NUM_HANDS:num_hands"
  input_side_packet: "MODEL_COMPLEXITY:model_complexity"
  input_side_packet: "USE_PREV_LANDMARKS:use_prev_landmarks"
  output_stream: "LANDMARKS:multi_hand_landmarks"
  output_stream: "WORLD_LANDMARKS:multi_hand_world_landmarks"
  output_stream: "HANDEDNESS:multi_handedness"
  output_stream: "PALM_DETECTIONS:palm_detections"
  output_stream: "HAND_ROIS_FROM_LANDMARKS:hand_rects"
  output_stream: "HAND_ROIS_FROM_PALM_DETECTIONS:hand_rects_from_palm_detections"
}