func analyse(img *floatimage.FloatImg) (min, max, mean, variance float32) {
var sum float64
bounds := img.Bounds()
if img.Bounds().Empty() {
return 0, 0, 0, 0
}
sum = 0.0
min = img.Pix[0]
max = img.Pix[0]
for y := bounds.Min.Y + 1; y < bounds.Max.Y-1; y++ {
for x := bounds.Min.X + 1; x < bounds.Max.X-1; x++ {
value := img.AtF(x, y)[0]
if value < min {
min = value
}
if value > max {
max = value
}
sum += float64(value)
}
}
imgsize := float64(bounds.Dx()-2) * float64(bounds.Dy()-2)
mean = float32(sum / imgsize)
variance = 0.0
for y := bounds.Min.Y + 1; y < bounds.Max.Y-1; y++ {
for x := bounds.Min.X + 1; x < bounds.Max.X-1; x++ {
temp := img.AtF(x, y)[0] - mean
variance += temp * temp
}
}
variance = float32(float64(variance) / imgsize)
return
}
// MagImage generates a magnitude image from an optic flow
// field and returns it as a single channel floatimage.FloatImg
func MagImage(uv *floatimage.FloatImg) (magImg *floatimage.FloatImg) {
bounds := uv.Bounds()
// Magnitude image
magImg = floatimage.NewFloatImg(bounds, 1)
// Calculate
for j := bounds.Min.Y; j < bounds.Max.Y; j++ {
for i := bounds.Min.X; i < bounds.Max.X; i++ {
vec := uv.AtF(i, j)
tmp := vec[0]*vec[0] + vec[1]*vec[1]
magImg.Set(i, j, 0, float32(math.Sqrt(float64(tmp))))
}
}
return
}
// OpticFlowHornSchunk computes the optic flow between two images
// the images need to have Dummie borders (see floatimage.Dummies())
// applied.
// It returns the optic flow field as a 2 channel floatimage.FloatImg
func OpticFlowHornSchunk(f1, f2 *floatimage.FloatImg, alpha float32, iterations int) (uv *floatimage.FloatImg) {
// Compute fx, fy, fz derivatives as FloatImg with 3 channels for faster access
derivs := deriveMixed(f1, f2)
// bounds without dummies
bounds := f1.Bounds()
// vector field as FloatImg with 2 channels
uv = floatimage.NewFloatImg(bounds, 2)
// temporary storage for vector field from previous iteration
uvOld := floatimage.NewFloatImg(bounds, 2)
// Process image using the Jacobi method to incrementally compute the vector field
for k := 1; k <= iterations; k++ {
flow(float32(alpha), derivs, uvOld, uv)
uvOld.Copy(uv)
}
return
}
func flow(alpha float32, derivs, oldvec, vecField *floatimage.FloatImg) {
bounds := vecField.Bounds()
help := 1.0 / alpha
var nn int
var uSum, vSum float32
var uv []float32
for j := bounds.Min.Y; j < bounds.Max.Y; j++ {
for i := bounds.Min.X; i < bounds.Max.X; i++ {
nn = 0
uSum, vSum = 0, 0
if i > bounds.Min.X {
nn++
uv = oldvec.AtF(i-1, j)
uSum += uv[0]
vSum += uv[1]
}
if i < bounds.Max.X-1 {
nn++
uv = oldvec.AtF(i+1, j)
uSum += uv[0]
vSum += uv[1]
}
if j > bounds.Min.Y {
nn++
uv = oldvec.AtF(i, j-1)
uSum += uv[0]
vSum += uv[1]
}
if j < bounds.Max.Y-1 {
nn++
uv = oldvec.AtF(i, j+1)
uSum += uv[0]
vSum += uv[1]
}
dvs := derivs.AtF(i, j)
fxij, fyij, fzij := dvs[Fxc], dvs[Fyc], dvs[Fzc]
uv = oldvec.AtF(i, j)
uSum -= help * fxij * (fyij*uv[1] + fzij)
uSum /= float32(nn) + help*fxij*fxij
vSum -= help * fyij * (fxij*uv[0] + fzij)
vSum /= float32(nn) + help*fyij*fyij
uv = vecField.AtF(i, j)
uv[0], uv[1] = uSum, vSum
}
}
}
func deriveMixed(f1, f2 *floatimage.FloatImg) *floatimage.FloatImg {
const hx = 1.0
const hy = 1.0
bounds := f1.Bounds()
derivs := floatimage.NewFloatImg(bounds, 3)
for j := bounds.Min.Y + 1; j < bounds.Max.Y-1; j++ {
for i := bounds.Min.X + 1; i < bounds.Max.X-1; i++ {
Fx := (f1.AtF(i+1, j)[0] - f1.AtF(i-1, j)[0] + f2.AtF(i+1, j)[0] - f2.AtF(i-1, j)[0]) / (4.0 * hx)
Fy := (f1.AtF(i, j+1)[0] - f1.AtF(i, j-1)[0] + f2.AtF(i, j+1)[0] - f2.AtF(i, j-1)[0]) / (4.0 * hy)
Fz := f2.AtF(i, j)[0] - f1.AtF(i, j)[0]
dvs := derivs.AtF(i, j)
dvs[Fxc], dvs[Fyc], dvs[Fzc] = Fx, Fy, Fz
}
}
return derivs
}