hollance / blazeface-pytorch Goto Github PK

weighted = (coordinates * scores).sum(dim=0) / total_score

coordinates * scores error:
RuntimeError: The size of tensor a (4) must match the size of tensor b (3) at non-singleton dimension 1

the loop

        while len(remaining) > 0:
            detection = detections[remaining[0]]

            # Compute the overlap between the first box and the other 
            # remaining boxes. (Note that the other_boxes also include
            # the first_box.)
            first_box = detection[:4]
            other_boxes = detections[remaining, :4]
            ious = overlap_similarity(first_box, other_boxes)

            # If two detections don't overlap enough, they are considered
            # to be from different faces.
            mask = ious > self.min_suppression_threshold
            overlapping = remaining[mask]
            remaining = remaining[~mask]

            # Take an average of the coordinates from the overlapping
            # detections, weighted by their confidence scores.
            weighted_detection = detection.clone()
            if len(overlapping) > 1:
                coordinates = detections[overlapping, :16]
                scores = detections[overlapping, 16:17]
                total_score = scores.sum()
                weighted = (coordinates * scores).sum(dim=0) / total_score
                weighted_detection[:16] = weighted
                weighted_detection[16] = total_score / len(overlapping)

            output_detections.append(weighted_detection)

is infinite if len(remaining) == 1. I faced this when remaining = tensor([1]). Should it not be while len(remaining) > 1. Note this case occured when i set the min_score and min_suppression thresholds to 0.172, 0.104 respectively (if thats of any use)

Will training process opensource for BlazeFace?

Thank you for your work. I could not find testing code. Could you please advice? Thank you.

Hello,

I am unable to load weights.

net.load_weights("blazeface.pth")

I am getting the below error.

---------------------------------------------------------------------------
ModuleAttributeError                      Traceback (most recent call last)
<ipython-input-29-cdf78203f6d0> in <module>
----> 1 net.load_weights("blazeface.pth")
      2 net.load_anchors("anchors.npy")
      3 
      4 # Optionally change the thresholds:
      5 net.min_score_thresh = 0.75

~/plaground/virtualenv/awslocal/lib/python3.7/site-packages/torch/nn/modules/module.py in __getattr__(self, name)
    770                 return modules[name]
    771         raise ModuleAttributeError("'{}' object has no attribute '{}'".format(
--> 772             type(self).__name__, name))
    773 
    774     def __setattr__(self, name: str, value: Union[Tensor, 'Module']) -> None:

ModuleAttributeError: 'BlazeFace' object has no attribute 'load_weights'

Hello,
Thanks to share this great job !
Just a question by curiosity: did you test that converted pytorch version provides similar results as the original TFLite one ?
:-)

I'm a beginner. I want to ask：How to use the blazeface.pth?

Hi! I'm currently starting some research about using face-detection models in limited compute devices such as smartphones (specifically I'm focusing on PyTorch Mobile) and my goal is to replicate the same app that Mediapipe did on Android but using your implementation of BlazeFace on PyTorch instead of TFLite.

Is this feasible?

hi,
I load the anchorsback.npy and the anchors.npy and print them,
find the anchors are same.

is it right??

I may be reading it wrong, but are the lines used to calculate max_xy using the box dimensions, when they should be using the boxes max x/y coordinates?

A box is defined as [min_x, min_y, dim_x, dim_y], is it not?

In Inference.ipynb In[5], """net.load_weights("blazeface.pth")""" goes wrong.I check the source code,torch receive a file object but not a string object,and I change just like:
"""
f = open("blazeface.pth", 'rb')
net.load_weights(f)
""".
After change, it can work properly.

There seems to be some issue, that this doesn't load, others are loading.

1. I saw the RAW content of this ipynb and saw few images in it - not sure whether they are the cause.

2. I had few questions which this code might have answered but I am not able to figure them out.
   a. the input to this network is 128x128 - I saw your comment in the raw code where it says we have to resize, if an image is available as 320x240 if we resize at will - the faces will be squashed - does the detector work with this or does it expect aspect ratios to be maintained.

b. How is the inference done after the concantenation layer, which produces a 1x896x1 output. What kind of detector is used?

c. You have experience with converting weights from tensorflow lite to pytorch. I have a requirement to transfer a model from pytorch to tensorflow lite, can you point me to a starter guide for this.

Does you conversion output the depth values of the xy image coordinates that is traditionally in the blazepose/blazeface models? Have you ever converted the full blazepose (full body landmark) models to pytorch? If not what modifications would need to be made to your conversion script if any?

Hello. Thanks for your work.
I see that there is still a blazeface_back model at https://github.com/google/mediapipe/blob/master/mediapipe/models/face_detection_back.tflite, how should i implement ?

Thanks very much for making this available.

I've been reading the code, and I think there might be a typo here: https://github.com/hollance/BlazeFace-PyTorch/blob/master/blazeface.py#L370

    max_xy = torch.min(box_a[:, 2:].unsqueeze(1).expand(A, B, 2),
                       box_b[:, 2:].unsqueeze(0).expand(A, B, 2))
    min_xy = torch.max(box_a[:, :2].unsqueeze(1).expand(A, B, 2),
                       box_b[:, :2].unsqueeze(0).expand(A, B, 2))

max_xy and min_xy are just reversed, right? Or am I missing something much smarter than me?

I don't think it affects anything either way though?