Transfer learning using MediaPipe model maker for BlazeFace Face Detector

Google AI Edge
1

Hi All,
From this table my understanding for MediaPipe face detect model ( https://ai.google.dev/edge/mediapipe/solutions/guide#available_solutions) Transfer Learning can not be used using model maker. Is my understanding correct? Can someone also help me understand why transfer learning is enable for Face detection model.
Also, I am planning to use https://ai.google.dev/edge/mediapipe/solutions/vision/face_landmarker and pose landmark detection in my application.

Is there any other easier solution similar to transfer learning for these models?

2

Any comment/suggestion?

3

you’re right — Model Maker doesn’t support transfer learning for BlazeFace / MediaPipe face detection, and there’s no official workaround for it.

reason is simple: that detector is a fixed, optimized MediaPipe graph (anchors + loss + post-processing tightly coupled), so it’s not exposed for fine-tuning like classification models.

for face_landmarker / pose, same story — they’re prebuilt Tasks models, not meant for retraining.

if you need customization, you either:

  • use them as-is and build logic on top, or

  • switch to a custom TFLite detector (SSD/MobileNet style) if you want real training control.

4

Thank you for the response @Linda_Anderson
I am thinking of retrieving SavedModel from BlazeFace tflite file, unfreeze top layer, perform transfer learning and then convert to tflite for running on edge devices. Do you see any problem with this approach? If no, then, whether this option (getting back to SavedModel) is better or I should take some other models like SSD/MobileNet kind of model. Also, help me understand why to go with specific approach?
I was planning to use MediaPipe models for face detection, landmark as well as for iris, mainly because these models are optimized for embedded devices.

5

Hi All, Please comment/suggest?