BlazeFace - How to use? about pinto_model_zoo

[PINTO0309]

Just concat each layer. My goal is to quantize it while maintaining accuracy, so I deliberately exclude the final layer of concat.

PINTO_model_zoo/030_BlazeFace/01_float32/08_new_blazeface_back_tflite2h5_weight_int_fullint_float16_quant.py

Lines 370 to 392 in dbc91cd

	# Block_33
	conv33_1 = Conv2D(filters=2, kernel_size=[1, 1], strides=[1, 1], padding="same", dilation_rate=[1, 1],
	kernel_initializer=Constant(np.load('weights_back/classificator_16_Kernel').transpose(1,2,3,0)),
	bias_initializer=Constant(np.load('weights_back/classificator_16_Bias')))(relu31_1)
	reshape33_1 = tf.reshape(conv33_1, (1, 512, 1), name='classificators_1')
	# Block_34
	conv34_1 = Conv2D(filters=6, kernel_size=[1, 1], strides=[1, 1], padding="same", dilation_rate=[1, 1],
	kernel_initializer=Constant(np.load('weights_back/classificator_32_Kernel').transpose(1,2,3,0)),
	bias_initializer=Constant(np.load('weights_back/classificator_32_Bias')))(relu32_1)
	reshape34_1 = tf.reshape(conv34_1, (1, 384, 1), name='classificators_2')
	# Block_35
	conv35_1 = Conv2D(filters=32, kernel_size=[1, 1], strides=[1, 1], padding="same", dilation_rate=[1, 1],
	kernel_initializer=Constant(np.load('weights_back/regressor_16_Kernel').transpose(1,2,3,0)),
	bias_initializer=Constant(np.load('weights_back/regressor_16_Bias')))(relu31_1)
	reshape35_1 = tf.reshape(conv35_1, (1, 512, 16), name='regressors_1')
	# Block_36
	conv36_1 = Conv2D(filters=96, kernel_size=[1, 1], strides=[1, 1], padding="same", dilation_rate=[1, 1],
	kernel_initializer=Constant(np.load('weights_back/regressor_32_Kernel').transpose(1,2,3,0)),
	bias_initializer=Constant(np.load('weights_back/regressor_32_Bias')))(relu32_1)
	reshape36_1 = tf.reshape(conv36_1, (1, 384, 16), name='regressors_2')

[vladmandic]

thanks, that (sort of) worked, just had to be careful with order of concat operations.

but...using same code for front and back model (just changing image inputsize) and it seems back model is far worse than front one - i expected that it would be able to handle smaller faces as advertised by mediapipe, but it only detects faces when it fills 50% of input frame - or i'm doing something wrong (could be issue with predefined anchors?).

do you have any experience with it?

[vladmandic]

nvm, it was a problem with my math on generating anchors - blazeface-back works just fine!
i'm closing this issue - thanks for your help!