GPU for pose detection enabled but no performance increase #6041
Description
Have I written custom code (as opposed to using a stock example script provided in MediaPipe)
Yes
OS Platform and Distribution
MacOS 15.5
MediaPipe Tasks SDK version
Version: 0.10.21
Task name (e.g. Image classification, Gesture recognition etc.)
Pose landmark detection (GPU enable)
Programming Language and version (e.g. C++, Python, Java)
Python
Describe the actual behavior
Framerate for pose detection is the same for CPU and GPU
Describe the expected behaviour
Framerate for pose detection is faster with GPU enabled.
Standalone code/steps you may have used to try to get what you need
import mediapipe as mp
import cv2
import time
from collections import deque
BaseOptions = mp.tasks.BaseOptions
PoseLandmarker = mp.tasks.vision.PoseLandmarker
PoseLandmarkerOptions = mp.tasks.vision.PoseLandmarkerOptions
PoseLandmarkerResult = mp.tasks.vision.PoseLandmarkerResult
VisionRunningMode = mp.tasks.vision.RunningMode
model_path = '/absolute/path/to/pose_landmarker.task'
class FPSMeter:
def __init__(self, window_seconds = 30.0):
self.timestamps = deque()
self.window = window_seconds
def update(self):
now = time.perf_counter()
self.timestamps.append(now)
while self.timestamps and (now-self.timestamps[0])>self.window:
self.timestamps.popleft()
if len(self.timestamps)>1:
duration = self.timestamps[-1] - self.timestamps[0]
fps = len(self.timestamps)/duration if duration > 0 else 0.0
else:
fps = 0.0
print(f"Average FPS (last {self.window:.0f}s): {fps:.2f}")
return fps
fps_meter = FPSMeter(window_seconds=30)
def print_result(result: PoseLandmarkerResult, output_image: mp.Image, timestamp_ms: int):
if result is None:
print('no result!')
else:
fps_meter.update()
options = PoseLandmarkerOptions(
base_options=BaseOptions(model_asset_path=model_path ,delegate=BaseOptions.Delegate.GPU),
running_mode=VisionRunningMode.LIVE_STREAM,
result_callback=print_result)
stream = cv2.VideoCapture(0)
if not stream.isOpened():
print("no stream :(")
exit()
#create the landmarker
landmarker = PoseLandmarker.create_from_options(options)
while (True):
ret, frame = stream.read()
frame_timestamp_ns = time.monotonic_ns()//1000
if not ret:
print("No more stream :( ")
break
mp_frame = mp.Image(image_format=mp.ImageFormat.SRGBA, data=cv2.cvtColor(frame, cv2.COLOR_BGR2RGBA))
landmarker.detect_async(mp_frame, frame_timestamp_ns)
if cv2.waitKey(1) == ord('q'):
break
stream.release() Other info / Complete Logs
When running with GPU in the base options:
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1752594257.192764 49864 gl_context.cc:369] GL version: 2.1 (2.1 Metal - 89.4), renderer: Apple M1 Pro
INFO: Created TensorFlow Lite delegate for Metal.
W0000 00:00:1752594257.332836 49967 landmark_projection_calculator.cc:186] Using NORM_RECT without IMAGE_DIMENSIONS is only supported for the square ROI. Provide IMAGE_DIMENSIONS or use PROJECTION_MATRIX.
... fps averages out over 30 seconds....
Average FPS (last 30s): 27.14When running with default:
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1752594148.404747 47815 gl_context.cc:369] GL version: 2.1 (2.1 Metal - 89.4), renderer: Apple M1 Pro
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
W0000 00:00:1752594148.459899 47936 inference_feedback_manager.cc:114] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
W0000 00:00:1752594148.468793 47935 inference_feedback_manager.cc:114] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
W0000 00:00:1752594148.515939 47940 landmark_projection_calculator.cc:186] Using NORM_RECT without IMAGE_DIMENSIONS is only supported for the square ROI. Provide IMAGE_DIMENSIONS or use PROJECTION_MATRIX
.... fps averages out over 30 seconds ....
Average FPS (last 30s): 27.14