convert_models.sh palm_detection_lite.tflite result blob file not work

[personball]

Hi @geaxgx , Thanks For all your work to convert mediapipe handtracker flow to OAK device, This saved a lot of time for me.

I'm working on depthai unity with OAK device, and had make palm detection and hand tracker success within your help(python code in this repository, By the way, I'm impl these with cpp).

My project is designed to track most like 10 hands.
It seems mediapipe palm_detection_lite mode in this issue PINTO0309/tflite2tensorflow#19 you mentioned has better accuracy, I'm success to convert it into a blob model follow your steps in mediapipe-models at readme, but it not work. didn't detect my hands.

Here is log file of convert_models.sh
convert.zip

the palm_detection_lite.tflite file is from mediapipe/third_party/external_files.bzl

    http_file(
        name = "com_google_mediapipe_palm_detection_lite_tflite",
        sha256 = "e9a4aaddf90dda56a87235303cf00e4c2d3fb28725f68fd88772997dac905c18",
        urls = ["https://storage.googleapis.com/mediapipe-assets/palm_detection_lite.tflite?generation=1661875885885770"],
    )

and has set pd_input_length from 128 to 192, also check anchors generate options, and decode_bboxes's scale.

So, maybe there is something I missed?