Retinaface Mobilenet V1 output

General
22

Sorry might be a bit late to the party but I was running into very similar issues and this post from another thread helped me out immensely: Help with inference/post-processing for the scrfd model for face detection - #3 by olof.

The issue that I was running into that the post unveiled was that the class layer output needed to be adjusted accordingly. Hope this helps.

def rescale_network_outputs(self, outputs):
        box_layer_names = ["scrfd_2_5g/conv43", "scrfd_2_5g/conv50", "scrfd_2_5g/conv56"]
        class_layer_names = ["scrfd_2_5g/conv42", "scrfd_2_5g/conv49", "scrfd_2_5g/conv55"]
        landmark_layer_names = ["scrfd_2_5g/conv44", "scrfd_2_5g/conv51", "scrfd_2_5g/conv57"]

        rescaled_outputs = []
        for output_name, output in outputs[0].items():
            # Convert to float32 to avoid overflow
            output = output.astype(np.float32)

            if output_name in box_layer_names:
                downscale_factor = 32 # Magic number, but probably the maximum downscale factor
                output = output / downscale_factor
            elif output_name in class_layer_names:
                # From range UINT8 [0, 255] to FLOAT32 [0, 1]
                output = output / 255
            elif output_name in landmark_layer_names:
                # Converts from Qunatized UINT8 to FLOAT32
                # These are approximate values as I couldn't find the exact values in the model
                # Exact values are determined when they compile from onnx to hef
                zero_point = 113
                scale = 29
                output = (output - zero_point) / scale
            else:
                raise ValueError(f"Unknown output name: {output_name}")
            
            reshaped = output.reshape(1, -1, output.shape[-1])
            rescaled_outputs.append(reshaped)

        return rescaled_outputs