You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We find this value (of 0.5f) to be too permissive, as its easy to "trick" the model into detecting random hands waving in front of the camera as a person. Using the Web SDK of blaze pose with a min_pose_detection_confidence of > 0.7f seems to work better with way fewer false positives.
Describe the solution you'd like
It would be ideal for this blaze pose param to be exposed in the Blazepose.Options.PoseLandmarkDetect.Options as a public float.
Describe alternatives you've considered
We tried playing with the Score threshold that is part of the Options, but it only affects the drawing rather than the detection itself. Also, we tried to change the "Score" filter on the landmarks (currently hardcoded to filter out <0.2. However, even bad quality inputs are being delivered with scores above 0.7, so its hard to to use this method to segregate bad and good quality inputs.
Additional context
Here is example of "min_pose_detection_confidence" scores set at default 0.5 and then at 0.8. With lower score, its easy to trick the detection model into thinking an arm is a person laying horizontally, with fingers being detected as limbs and wrist as shoulders. With score of 0.8, we can filter those cases out.
The text was updated successfully, but these errors were encountered:
Hi @asus4, to also add to the above, @Ale1 and I tried that on multiple devices (iPad 2021 12.9, iPhone 12 Pro, Samsung Galaxy S10 to name a few) and this occurs on all.
Ale1
changed the title
Expose blaze pose config "min_pose_detection_confidence"
Expose mediapipe config option "min_pose_detection_confidence"
Jun 11, 2023
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.
Is your feature request related to a problem? Please describe.
The pose landmark detection model seems to be using the default value of the param "min_pose_detection_confidence". The effects of this value can be seen here: https://mediapipe-studio.webapps.google.com/demo/pose_landmarker. And the documentation for this param here: https://developers.google.com/mediapipe/solutions/vision/pose_landmarker#configurations_options.
We find this value (of 0.5f) to be too permissive, as its easy to "trick" the model into detecting random hands waving in front of the camera as a person. Using the Web SDK of blaze pose with a min_pose_detection_confidence of > 0.7f seems to work better with way fewer false positives.
Describe the solution you'd like
It would be ideal for this blaze pose param to be exposed in the Blazepose.Options.PoseLandmarkDetect.Options as a public float.
Describe alternatives you've considered
We tried playing with the Score threshold that is part of the Options, but it only affects the drawing rather than the detection itself. Also, we tried to change the "Score" filter on the landmarks (currently hardcoded to filter out <0.2. However, even bad quality inputs are being delivered with scores above 0.7, so its hard to to use this method to segregate bad and good quality inputs.
Additional context
Here is example of "min_pose_detection_confidence" scores set at default 0.5 and then at 0.8. With lower score, its easy to trick the detection model into thinking an arm is a person laying horizontally, with fingers being detected as limbs and wrist as shoulders. With score of 0.8, we can filter those cases out.
The text was updated successfully, but these errors were encountered: