pico：200 行实现的面部识别库
OneFile
源码
/* This library is released under the MIT license, see https://github.com/nenadmarkus/picojs */
pico = {}
pico.unpack_cascade = function(bytes)
{
//
const dview = new DataView(new ArrayBuffer(4));
/*
we skip the first 8 bytes of the cascade file
(cascade version number and some data used during the learning process)
*/
let p = 8;
/*
read the depth (size) of each tree first: a 32-bit signed integer
*/
dview.setUint8(0, bytes[p+0]), dview.setUint8(1, bytes[p+1]), dview.setUint8(2, bytes[p+2]), dview.setUint8(3, bytes[p+3]);
const tdepth = dview.getInt32(0, true);
p = p + 4
/*
next, read the number of trees in the cascade: another 32-bit signed integer
*/
dview.setUint8(0, bytes[p+0]), dview.setUint8(1, bytes[p+1]), dview.setUint8(2, bytes[p+2]), dview.setUint8(3, bytes[p+3]);
const ntrees = dview.getInt32(0, true);
p = p + 4
/*
read the actual trees and cascade thresholds
*/
const tcodes_ls = [];
const tpreds_ls = [];
const thresh_ls = [];
for(let t=0; t<ntrees; ++t)
{
// read the binary tests placed in internal tree nodes
Array.prototype.push.apply(tcodes_ls, [0, 0, 0, 0]);
Array.prototype.push.apply(tcodes_ls, bytes.slice(p, p+4*Math.pow(2, tdepth)-4));
p = p + 4*Math.pow(2, tdepth)-4;
// read the prediction in the leaf nodes of the tree
for(let i=0; i<Math.pow(2, tdepth); ++i)
{
dview.setUint8(0, bytes[p+0]), dview.setUint8(1, bytes[p+1]), dview.setUint8(2, bytes[p+2]), dview.setUint8(3, bytes[p+3]);
tpreds_ls.push(dview.getFloat32(0, true));
p = p + 4;
}
// read the threshold
dview.setUint8(0, bytes[p+0]), dview.setUint8(1, bytes[p+1]), dview.setUint8(2, bytes[p+2]), dview.setUint8(3, bytes[p+3]);
thresh_ls.push(dview.getFloat32(0, true));
p = p + 4;
}
const tcodes = new Int8Array(tcodes_ls);
const tpreds = new Float32Array(tpreds_ls);
const thresh = new Float32Array(thresh_ls);
/*
construct the classification function from the read data
*/
function classify_region(r, c, s, pixels, ldim)
{
r = 256*r;
c = 256*c;
let root = 0;
let o = 0.0;
const pow2tdepth = Math.pow(2, tdepth) >> 0; // '>>0' transforms this number to int
for(let i=0; i<ntrees; ++i)
{
let idx = 1;
for(let j=0; j<tdepth; ++j)
// we use '>> 8' here to perform an integer division: this seems important for performance
idx = 2*idx + (pixels[((r+tcodes[root + 4*idx + 0]*s) >> 8)*ldim+((c+tcodes[root + 4*idx + 1]*s) >> 8)]<=pixels[((r+tcodes[root + 4*idx + 2]*s) >> 8)*ldim+((c+tcodes[root + 4*idx + 3]*s) >> 8)]);
o = o + tpreds[pow2tdepth*i + idx-pow2tdepth];
if(o<=thresh[i])
return -1;
root += 4*pow2tdepth;
}
return o - thresh[ntrees-1];
}
/*
we're done
*/
return classify_region;
}
pico.run_cascade = function(image, classify_region, params)
{
const pixels = image.pixels;
const nrows = image.nrows;
const ncols = image.ncols;
const ldim = image.ldim;
const shiftfactor = params.shiftfactor;
const minsize = params.minsize;
const maxsize = params.maxsize;
const scalefactor = params.scalefactor;
let scale = minsize;
const detections = [];
while(scale<=maxsize)
{
const step = Math.max(shiftfactor*scale, 1) >> 0; // '>>0' transforms this number to int
const offset = (scale/2 + 1) >> 0;
for(let r=offset; r<=nrows-offset; r+=step)
for(let c=offset; c<=ncols-offset; c+=step)
{
const q = classify_region(r, c, scale, pixels, ldim);
if (q > 0.0)
detections.push([r, c, scale, q]);
}
scale = scale*scalefactor;
}
return detections;
}
pico.cluster_detections = function(dets, iouthreshold)
{
/*
sort detections by their score
*/
dets = dets.sort(function(a, b) {
return b[3] - a[3];
});
/*
this helper function calculates the intersection over union for two detections
*/
function calculate_iou(det1, det2)
{
// unpack the position and size of each detection
const r1=det1[0], c1=det1[1], s1=det1[2];
const r2=det2[0], c2=det2[1], s2=det2[2];
// calculate detection overlap in each dimension
const overr = Math.max(0, Math.min(r1+s1/2, r2+s2/2) - Math.max(r1-s1/2, r2-s2/2));
const overc = Math.max(0, Math.min(c1+s1/2, c2+s2/2) - Math.max(c1-s1/2, c2-s2/2));
// calculate and return IoU
return overr*overc/(s1*s1+s2*s2-overr*overc);
}
/*
do clustering through non-maximum suppression
*/
const assignments = new Array(dets.length).fill(0);
const clusters = [];
for(let i=0; i<dets.length; ++i)
{
// is this detection assigned to a cluster?
if(assignments[i]==0)
{
// it is not:
// now we make a cluster out of it and see whether some other detections belong to it
let r=0.0, c=0.0, s=0.0, q=0.0, n=0;
for(let j=i; j<dets.length; ++j)
if(calculate_iou(dets[i], dets[j])>iouthreshold)
{
assignments[j] = 1;
r = r + dets[j][0];
c = c + dets[j][1];
s = s + dets[j][2];
q = q + dets[j][3];
n = n + 1;
}
// make a cluster representative
clusters.push([r/n, c/n, s/n, q]);
}
}
return clusters;
}
pico.instantiate_detection_memory = function(size)
{
/*
initialize a circular buffer of `size` elements
*/
let n = 0;
const memory = [];
for(let i=0; i<size; ++i)
memory.push([]);
/*
build a function that:
(1) inserts the current frame's detections into the buffer;
(2) merges all detections from the last `size` frames and returns them
*/
function update_memory(dets)
{
memory[n] = dets;
n = (n+1)%memory.length;
dets = [];
for(i=0; i<memory.length; ++i)
dets = dets.concat(memory[i]);
//
return dets;
}
/*
we're done
*/
return update_memory;
}