stanislasbertrand / retinaface-tf2 Goto Github PK

Hi, so im trying to get Region of Interest from detected faces and then blur them using this code:

for i in range(faces.shape[0]):
box = faces[i].astype(np.int)
face = image[(box[1], box[3]), (box[0], box[2])]
face = anonymize_face_pixelate(face, blocks=3)
image[(box[1], box[3]), (box[0], box[2])] = face

I have my function "anonymize_face_pixelate" that will pixelate that region of interest but something is wrong probably with format of coordinates in "box[-]". When i run code, it executes without error, even print correct number of detected faces, but they are not pixelated. Any advice? Thank you so much

Hi,

I noticed that files requirements.txt and requirements_gpu.txt have the same content:

tensorflow-gpu==2.1.0
opencv-python
cython

Is there something I am missing or could requirements_gpu.txt just be removed since TF2 has only one Python dependency for both CPU and GPU ?

Thank you for your answer

Hey there, I saw your RetinaFace facedetector and Im trying to test it as you say in your github project but when i try run it, i can see red underline in bbox_transform.py because its not able to import that bbox_overlaps_cython from bbox.pyx.I installed all neccesary libraries and packages icluding cython but still not working. Is there any chance that you can help me how to run this thing? Im using Pycharm and python 3.6 on iMac OSX
Thank you so much
Sincerelly Peter

In this line img_w, img_h = img.shape[0:2], but image shape should be [h, w] rather than [w, h]

I use the https://github.com/biubug6/Pytorch_Retinaface to validate widerface and the resuslt is strange.

For example,I only test one image. The Image grouth only have one face. I use the network to detect the image and the result have two face be detect. One of the two is incorrect.
When I compute the AP, and the AP is 1. Is it correct ?

and why operating the image tensor as following, in RetinaFace.detect():

for i in range(3):
        im_tensor[0, i, :, :] = (img[:, :, 2 - i] / pixel_scale - pixel_means[2 - i]) / pixel_stds[
            2 - i]
net_out = self.model.predict(im_tensor.transpose(0, 2, 3, 1))

Thanks in advance!

How can we convert this model into .h5 or savedmodel or tflite?

Hi, I am trying to use this in a windows enviroment. Running the Makefile with GnuWin32 only show the commands in the makefile. Running the commands manually in the rcnn folders specified returns:

AttributeError: 'MSVCCompiler' object has no attribute 'compiler_so'

Hi. Thanks for providing TF implementation!
Which backbone arch (Resnet/mobilenet...) does the pre-trained model belong to?

I don't understand the third step in the installation process: 'Then build the RCNN Module by running: make'

What should I compile.

faces, landmarks = detector.detect(img, 0.9)

/RetinaFace-tf2/src/retinafacetf2/networks/network_utils.py in load_weights(weight_file)
---> 16         weights_dict = np.load(weight_file, allow_pickle=True, encoding='bytes').item()
OSError: Failed to interpret file './data/retinafaceweights.npy' as a pickle

os.path.exists('./data/retinafaceweights.npy')
True

both in 1.19 numpy and 1.22 numpy

And how long time it takes to train a model?
Do you train the model on a single GPU or multi GPUs?