I'm working on facenet with openvino R3 release on win10 cpu.
I converted the facenet models: 20180402-114759 , but the output embeddings is different with the original tensorflow.
my steps following with the post: [Facenet] different output matrix with R3: https://software.intel.com/en-us/forums/computer-vision/topic/797938 :
I think the output matrix should be the same(or similar) with the same image between tensorflow and IR model. but the dist_embedding is larger than 0.35 and validate it on LFW the accuracy is only 0.990. compared with tf version 0.995
Does anyone have the same problem
By LFW I am assuming you mean Labeled Faces In The Wild Database
When you do python mo_tf.py -h you will see options --scale and --mean_values. Make sure that these exactly match that of Tensorflow. In other words, the image resizing/scaling in Tensorflow and Inference Engine may be different. It would be better to first convert your LFW images to bmp (a lossless format) and then resize the image using your favorite image editing utility to the correct topology (i.e. TensorFlow ) input size. Then re-run inference with both TensorFlow and Inference Engine once again and compare output results.
It is necessary to investigate what kind of input image pre-processing is performed in the topology or before the image is passed to the Tensorflow. So my recommendation is take a look at the FaceNet git repo, analyze how the image is passed to the TensorFlow model and is there any pre-processing that is a part of the topology ? (during conversion we remove part of the topology by cutting it with the --input_tensor).