vww

VWW (Visual Wake Words) — Lightweight convolutional neural network library for human detection. Based on MobileNet-like depthwise separable convolution architecture, INT8 quantized, pure C99 implementation, zero external dependencies.

API

#include <vww.h>

int vision_vww(struct vision_t * v);

Parameter	Description
v	Input image (grayscale or RGB format, resolution must be >= 96x96), internally center-cropped (preserving aspect ratio, shorter side aligned to 96 pixels) then resized to 96x96. If the input resolution is < 96x96, use vision_resize to upscale first
Return value	Human confidence percentage (0-100)

Shell Command

The vww command is used to detect humans in an image.

vww <image>

# Detect humans in an image
vww person.jpg
# Output: person: 87%

Network Structure

MobileNet-like depthwise separable convolutional network, 12 layers total.

Layer	Type	Input	Output	Kernel	Stride
L0	CONV2D	96x96x3	48x48x8	3x3	2x2
L1	DWCONV2D	48x48x8	48x48x8	3x3	1x1
L2	CONV2D	48x48x8	48x48x16	1x1	1x1
L3	DWCONV2D	48x48x16	24x24x16	3x3	2x2
L4	CONV2D	24x24x16	24x24x32	1x1	1x1
L5	DWCONV2D	24x24x32	24x24x32	3x3	1x1
L6	CONV2D	24x24x32	24x24x32	1x1	1x1
L7	DWCONV2D	24x24x32	12x12x32	3x3	2x2
L8	CONV2D	12x12x32	12x12x64	1x1	1x1
L9	DWCONV2D	12x12x64	12x12x64	3x3	1x1
L10	CONV2D	12x12x64	12x12x64	1x1	1x1
L11	DWCONV2D	12x12x64	6x6x64	3x3	2x2

Output goes through global average pooling + fully connected layer for classification.

Quantization

• Weights: INT8
• Biases: INT32
• Scale factors: INT32
• Activation function: ReLU

All weights and biases are embedded in the code at compile time, no external model file required.