Example #1
0
// CorrMultiBankStrideBLAS computes the strided correlation of
// a multi-channel image with a bank of multi-channel filters.
// 	h_p[u, v] = sum_q (f_q corr g_pq)[stride*u, stride*v]
func CorrMultiBankStrideBLAS(f *rimg64.Multi, g *MultiBank, stride int) (*rimg64.Multi, error) {
	out := ValidSizeStride(f.Size(), g.Size(), stride)
	if out.X <= 0 || out.Y <= 0 {
		return nil, nil
	}
	h := rimg64.NewMulti(out.X, out.Y, len(g.Filters))
	// Size of filters.
	m, n, k := g.Width, g.Height, g.Channels
	// Express as dense matrix multiplication.
	//   h_p[u, v] = sum_q (f_q corr g_pq)[u, v]
	//   h = A(f) X(g)
	// where A is whk by mnk
	// with w = ceil[(M-m+1)/stride],
	//      h = ceil[(N-n+1)/stride].
	a := blas.NewMat(h.Width*h.Height, m*n*k)
	{
		var r int
		for u := 0; u < h.Width; u++ {
			for v := 0; v < h.Height; v++ {
				var s int
				for i := 0; i < g.Width; i++ {
					for j := 0; j < g.Height; j++ {
						for q := 0; q < g.Channels; q++ {
							a.Set(r, s, f.At(stride*u+i, stride*v+j, q))
							s++
						}
					}
				}
				r++
			}
		}
	}
	x := blas.NewMat(m*n*k, h.Channels)
	{
		var r int
		for i := 0; i < g.Width; i++ {
			for j := 0; j < g.Height; j++ {
				for q := 0; q < g.Channels; q++ {
					for p := 0; p < h.Channels; p++ {
						x.Set(r, p, g.Filters[p].At(i, j, q))
					}
					r++
				}
			}
		}
	}
	y := blas.MatMul(1, a, x)
	{
		var r int
		for u := 0; u < h.Width; u++ {
			for v := 0; v < h.Height; v++ {
				for p := 0; p < h.Channels; p++ {
					h.Set(u, v, p, y.At(r, p))
				}
				r++
			}
		}
	}
	return h, nil
}
Example #2
0
// CorrMultiStrideBLAS computes the strided correlation of
// a multi-channel image with a multi-channel filter.
// 	h[u, v] = sum_q (f_q corr g_q)[stride*u, stride*v]
func CorrMultiStrideBLAS(f, g *rimg64.Multi, stride int) (*rimg64.Image, error) {
	out := ValidSizeStride(f.Size(), g.Size(), stride)
	if out.X <= 0 || out.Y <= 0 {
		return nil, nil
	}
	h := rimg64.New(out.X, out.Y)
	// Size of filters.
	m, n, k := g.Width, g.Height, g.Channels
	// Express as dense matrix multiplication.
	//   h[u, v] = sum_q (f_q corr g_q)[stride*u, stride*v]
	//   y(h) = A(f) x(g)
	// where A is wh by mnk
	// with w = ceil[(M-m+1)/stride],
	//      h = ceil[(N-n+1)/stride].
	a := blas.NewMat(h.Width*h.Height, m*n*k)
	{
		var r int
		for u := 0; u < h.Width; u++ {
			for v := 0; v < h.Height; v++ {
				var s int
				for i := 0; i < g.Width; i++ {
					for j := 0; j < g.Height; j++ {
						for q := 0; q < g.Channels; q++ {
							a.Set(r, s, f.At(stride*u+i, stride*v+j, q))
							s++
						}
					}
				}
				r++
			}
		}
	}
	x := blas.NewMat(m*n*k, 1)
	{
		var r int
		for i := 0; i < g.Width; i++ {
			for j := 0; j < g.Height; j++ {
				for q := 0; q < g.Channels; q++ {
					x.Set(r, 0, g.At(i, j, q))
					r++
				}
			}
		}
	}
	y := blas.MatMul(1, a, x)
	{
		var r int
		for u := 0; u < h.Width; u++ {
			for v := 0; v < h.Height; v++ {
				h.Set(u, v, y.At(r, 0))
				r++
			}
		}
	}
	return h, nil
}
Example #3
0
// CorrMultiStrideFFT computes the correlation of
// a multi-channel image with a multi-channel filter.
// 	h[u, v] = sum_q (f_q corr g_q)[u, v]
func CorrMultiStrideFFT(f, g *rimg64.Multi, stride int) (*rimg64.Image, error) {
	if err := errIfChannelsNotEq(f, g); err != nil {
		panic(err)
	}
	out := ValidSizeStride(f.Size(), g.Size(), stride)
	if out.X <= 0 || out.Y <= 0 {
		return nil, nil
	}
	// Compute strided convolution as the sum over
	// a stride x stride grid of small convolutions.
	grid := image.Pt(stride, stride)
	// But do not divide into a larger grid than the size of the filter.
	// If the filter is smaller than the stride,
	// then some pixels in the image will not affect the output.
	grid.X = min(grid.X, g.Width)
	grid.Y = min(grid.Y, g.Height)
	// Determine the size of the sub-sampled filter.
	gsub := image.Pt(ceilDiv(g.Width, grid.X), ceilDiv(g.Height, grid.Y))
	// The sub-sampled size of the image should be such that
	// the output size is attained.
	fsub := image.Pt(out.X+gsub.X-1, out.Y+gsub.Y-1)

	// Determine optimal size for FFT.
	work, _ := FFT2Size(fsub)
	// Cache FFT of each channel of image for convolving with multiple filters.
	// Re-use plan for multiple convolutions too.
	fhat := fftw.NewArray2(work.X, work.Y)
	ffwd := fftw.NewPlan2(fhat, fhat, fftw.Forward, fftw.Estimate)
	defer ffwd.Destroy()
	ghat := fftw.NewArray2(work.X, work.Y)
	gfwd := fftw.NewPlan2(ghat, ghat, fftw.Forward, fftw.Estimate)
	defer gfwd.Destroy()
	// Normalization factor.
	alpha := complex(1/float64(work.X*work.Y), 0)
	// Add the convolutions over channels and strides.
	hhat := fftw.NewArray2(work.X, work.Y)
	for k := 0; k < f.Channels; k++ {
		for i := 0; i < grid.X; i++ {
			for j := 0; j < grid.Y; j++ {
				// Copy each downsampled channel and take its transform.
				copyChannelStrideTo(fhat, f, k, stride, image.Pt(i, j))
				ffwd.Execute()
				copyChannelStrideTo(ghat, g, k, stride, image.Pt(i, j))
				gfwd.Execute()
				addMul(hhat, ghat, fhat)
			}
		}
	}
	// Take the inverse transform.
	h := rimg64.New(out.X, out.Y)
	scale(alpha, hhat)
	fftw.IFFT2To(hhat, hhat)
	copyRealTo(h, hhat)
	return h, nil
}
Example #4
0
File: dec.go Project: jvlmdr/go-cv
// DecimateMulti takes every r-th sample starting at (0, 0).
func DecimateMulti(f *rimg64.Multi, r int) *rimg64.Multi {
	out := ceilDivPt(f.Size(), r)
	g := rimg64.NewMulti(out.X, out.Y, f.Channels)
	for i := 0; i < g.Width; i++ {
		for j := 0; j < g.Height; j++ {
			for k := 0; k < g.Channels; k++ {
				g.Set(i, j, k, f.At(r*i, r*j, k))
			}
		}
	}
	return g
}
Example #5
0
File: conv.go Project: jvlmdr/go-cv
func (phi *ConvEach) Apply(x *rimg64.Multi) (*rimg64.Multi, error) {
	channels := x.Channels * len(phi.Filters.Filters)
	field := image.Pt(phi.Filters.Width, phi.Filters.Height)
	size := slide.ValidSize(x.Size(), field)
	y := rimg64.NewMulti(size.X, size.Y, channels)
	var n int
	for i := 0; i < x.Channels; i++ {
		// Convolve each channel of the input with the bank.
		yi, err := slide.CorrBankBLAS(x.Channel(i), phi.Filters)
		if err != nil {
			return nil, err
		}
		for j := 0; j < yi.Channels; j++ {
			// Copy the channels into the output.
			y.SetChannel(n, yi.Channel(j))
			n++
		}
	}
	return y, nil
}
Example #6
0
// CorrMultiBankFFT computes the correlation of
// a multi-channel image with a bank of multi-channel filters.
// 	h_p[u, v] = sum_q (f_q corr g_pq)[u, v]
func CorrMultiBankFFT(f *rimg64.Multi, g *MultiBank) (*rimg64.Multi, error) {
	out := ValidSize(f.Size(), g.Size())
	if out.X <= 0 || out.Y <= 0 {
		return nil, nil
	}
	// Determine optimal size for FFT.
	work, _ := FFT2Size(f.Size())
	// Cache FFT of each channel of image.
	fhat := make([]*fftw.Array2, f.Channels)
	for i := range fhat {
		fhat[i] = fftw.NewArray2(work.X, work.Y)
		copyChannelTo(fhat[i], f, i)
		fftw.FFT2To(fhat[i], fhat[i])
	}

	curr := fftw.NewArray2(work.X, work.Y)
	fwd := fftw.NewPlan2(curr, curr, fftw.Forward, fftw.Estimate)
	defer fwd.Destroy()
	sum := fftw.NewArray2(work.X, work.Y)
	bwd := fftw.NewPlan2(sum, sum, fftw.Backward, fftw.Estimate)
	defer bwd.Destroy()

	h := rimg64.NewMulti(out.X, out.Y, len(g.Filters))
	alpha := complex(1/float64(work.X*work.Y), 0)
	// For each output channel.
	for p, gp := range g.Filters {
		zero(sum)
		// For each input channel.
		for q := 0; q < f.Channels; q++ {
			// Take FFT of this input channel.
			copyChannelTo(curr, gp, q)
			fwd.Execute()
			// h_p[x] = (G_qp corr F_p)[x]
			// H_p[x] = conj(G_qp[x]) F_p[x]
			addScaleMul(sum, alpha, curr, fhat[q])
		}
		bwd.Execute()
		copyRealToChannel(h, p, sum)
	}
	return h, nil
}
Example #7
0
// CorrMultiBankFFT computes the correlation of
// a multi-channel image with a multi-channel filter.
// 	h[u, v] = sum_p (f_p corr g_p)[u, v]
func CorrMultiFFT(f, g *rimg64.Multi) (*rimg64.Image, error) {
	if err := errIfChannelsNotEq(f, g); err != nil {
		panic(err)
	}
	out := ValidSize(f.Size(), g.Size())
	if out.Eq(image.ZP) {
		return nil, nil
	}
	work, _ := FFT2Size(f.Size())
	fhat := fftw.NewArray2(work.X, work.Y)
	ghat := fftw.NewArray2(work.X, work.Y)
	ffwd := fftw.NewPlan2(fhat, fhat, fftw.Forward, fftw.Estimate)
	defer ffwd.Destroy()
	gfwd := fftw.NewPlan2(ghat, ghat, fftw.Forward, fftw.Estimate)
	defer gfwd.Destroy()
	hhat := fftw.NewArray2(work.X, work.Y)
	for p := 0; p < f.Channels; p++ {
		// Take transform of each channel.
		copyChannelTo(fhat, f, p)
		ffwd.Execute()
		copyChannelTo(ghat, g, p)
		gfwd.Execute()
		addMul(hhat, ghat, fhat)
	}
	n := float64(work.X * work.Y)
	scale(complex(1/n, 0), hhat)
	fftw.IFFT2To(hhat, hhat)
	h := rimg64.New(out.X, out.Y)
	copyRealTo(h, hhat)
	return h, nil
}
Example #8
0
// CorrMultiStrideNaive computes the correlation of
// a multi-channel image with a multi-channel filter.
// 	h[u, v] = sum_q (f_q corr g_q)[u, v]
func CorrMultiStrideNaive(f, g *rimg64.Multi, stride int) (*rimg64.Image, error) {
	if err := errIfChannelsNotEq(f, g); err != nil {
		panic(err)
	}
	out := ValidSizeStride(f.Size(), g.Size(), stride)
	h := rimg64.New(out.X, out.Y)
	for i := 0; i < h.Width; i++ {
		for j := 0; j < h.Height; j++ {
			var total float64
			for u := 0; u < g.Width; u++ {
				for v := 0; v < g.Height; v++ {
					p := image.Pt(i, j).Mul(stride).Add(image.Pt(u, v))
					for k := 0; k < f.Channels; k++ {
						total += f.At(p.X, p.Y, k) * g.At(u, v, k)
					}
				}
			}
			h.Set(i, j, total)
		}
	}
	return h, nil
}
Example #9
0
// CorrMultiBankStrideNaive computes the strided correlation of
// a multi-channel image with a bank of multi-channel filters.
// 	h_p[u, v] = sum_q (f_q corr g_pq)[stride*u, stride*v]
func CorrMultiBankStrideNaive(f *rimg64.Multi, g *MultiBank, stride int) (*rimg64.Multi, error) {
	out := ValidSizeStride(f.Size(), g.Size(), stride)
	if out.X <= 0 || out.Y <= 0 {
		return nil, nil
	}
	h := rimg64.NewMulti(out.X, out.Y, len(g.Filters))
	for u := 0; u < h.Width; u++ {
		for v := 0; v < h.Height; v++ {
			for p := 0; p < h.Channels; p++ {
				for i := 0; i < g.Width; i++ {
					for j := 0; j < g.Height; j++ {
						for q := 0; q < g.Channels; q++ {
							val := f.At(stride*u+i, stride*v+j, q) * g.Filters[p].At(i, j, q)
							h.Set(u, v, p, h.At(u, v, p)+val)
						}
					}
				}
			}
		}
	}
	return h, nil
}
Example #10
0
func (f *AffineScorer) Score(x *rimg64.Multi) (float64, error) {
	if f.Op != Cos {
		panic("cosine unimplemented")
	}
	if !x.Size().Eq(f.Tmpl.Size()) {
		return 0, fmt.Errorf("different size: input %v, template %v", x.Size(), f.Tmpl.Size())
	}
	if x.Channels != f.Tmpl.Channels {
		return 0, fmt.Errorf("different channels: input %v, template %v", x.Channels, f.Tmpl.Channels)
	}
	size := f.Tmpl.Size()
	var y float64
	for i := 0; i < size.X; i++ {
		for j := 0; j < size.Y; j++ {
			for k := 0; k < f.Tmpl.Channels; k++ {
				y += x.At(i, j, k) * f.Tmpl.At(i, j, k)
			}
		}
	}
	y += f.Bias
	return y, nil
}
Example #11
0
// CorrMultiBankNaive computes the correlation of
// a multi-channel image with a bank of multi-channel filters.
// 	h_p[u, v] = sum_q (f_q corr g_pq)[u, v]
func CorrMultiBankNaive(f *rimg64.Multi, g *MultiBank) (*rimg64.Multi, error) {
	out := ValidSize(f.Size(), g.Size())
	if out.X <= 0 || out.Y <= 0 {
		return nil, nil
	}
	h := rimg64.NewMulti(out.X, out.Y, len(g.Filters))
	for u := 0; u < h.Width; u++ {
		for v := 0; v < h.Height; v++ {
			for p := 0; p < h.Channels; p++ {
				var sum float64
				for i := 0; i < g.Width; i++ {
					for j := 0; j < g.Height; j++ {
						for q := 0; q < g.Channels; q++ {
							sum += f.At(i+u, j+v, q) * g.Filters[p].At(i, j, q)
						}
					}
				}
				h.Set(u, v, p, sum)
			}
		}
	}
	return h, nil
}
Example #12
0
// CorrMultiNaive computes the correlation of
// a multi-channel image with a multi-channel filter.
// 	h[u, v] = sum_p (f_p corr g_p)[u, v]
func CorrMultiNaive(f, g *rimg64.Multi) (*rimg64.Image, error) {
	if err := errIfChannelsNotEq(f, g); err != nil {
		panic(err)
	}
	out := ValidSize(f.Size(), g.Size())
	if out.Eq(image.ZP) {
		return nil, nil
	}
	h := rimg64.New(out.X, out.Y)
	for i := 0; i < out.X; i++ {
		for j := 0; j < out.Y; j++ {
			var total float64
			for u := 0; u < g.Width; u++ {
				for v := 0; v < g.Height; v++ {
					for p := 0; p < f.Channels; p++ {
						total += f.At(i+u, j+v, p) * g.At(u, v, p)
					}
				}
			}
			h.Set(i, j, total)
		}
	}
	return h, nil
}
Example #13
0
func errIfNotEqMulti(f, g *rimg64.Multi, eps float64) error {
	if !f.Size().Eq(g.Size()) {
		return fmt.Errorf("different size: %v, %v", f.Size(), g.Size())
	}
	if f.Channels != g.Channels {
		return fmt.Errorf("different channels: %d, %d", f.Channels, g.Channels)
	}
	for i := 0; i < f.Width; i++ {
		for j := 0; j < f.Height; j++ {
			for k := 0; k < f.Channels; k++ {
				a, b := f.At(i, j, k), g.At(i, j, k)
				if math.Abs(a-b) > eps*math.Max(math.Abs(a), math.Abs(b)) {
					return fmt.Errorf("different at x %d, y %d, c %d: %g, %g", i, j, k, a, b)
				}
			}
		}
	}
	return nil
}
Example #14
0
// CorrMultiAuto computes the correlation of
// a multi-channel image with a multi-channel filter.
// 	h[u, v] = sum_p (f_p corr g_p)[u, v]
// Automatically selects between naive and Fourier-domain convolution.
func CorrMultiAuto(f, g *rimg64.Multi) (*rimg64.Image, error) {
	if err := errIfChannelsNotEq(f, g); err != nil {
		panic(err)
	}
	// Size of output.
	size := ValidSize(f.Size(), g.Size())
	// Return empty image if that's the result.
	if size.Eq(image.ZP) {
		return nil, nil
	}
	// Need to compute one inner product per output element.
	naiveMuls := size.X * size.Y * g.Width * g.Height
	// Optimal FFT size and number of multiplications.
	_, fftMuls := FFT2Size(f.Size())
	// Need to perform two forward and one inverse transform.
	fftMuls *= 3
	// Switch implementation based on image size.
	if fftMuls < naiveMuls {
		return CorrMultiFFT(f, g)
	}
	return CorrMultiNaive(f, g)
}
Example #15
0
// CorrMultiBankBLAS computes the correlation of
// a multi-channel image with a bank of multi-channel filters.
// 	h_p[u, v] = sum_q (f_q corr g_pq)[u, v]
func CorrMultiBankBLAS(f *rimg64.Multi, g *MultiBank) (*rimg64.Multi, error) {
	out := ValidSize(f.Size(), g.Size())
	if out.X <= 0 || out.Y <= 0 {
		return nil, nil
	}
	// Express as dense matrix multiplication.
	//   h_p[u, v] = sum_q (f_q corr g_pq)[u, v]
	//   Y(h) = A(f) X(g)
	// If the number of input and output channels are Q and P, then
	//   A is (M-m+1)(N-n+1) x mnQ and
	//   X is mnQ x P, so that
	//   Y is (M-m+1)(N-n+1) x P.
	// Note that the time to build the system is therefore
	// affected more by the number of input channels Q than outputs P.

	h := rimg64.NewMulti(out.X, out.Y, len(g.Filters))
	M, N, K := h.Width, h.Height, h.Channels
	m, n, k := g.Width, g.Height, g.Channels
	a := blas.NewMat(M*N, m*n*k)
	{
		var r int
		for u := 0; u < h.Width; u++ {
			for v := 0; v < h.Height; v++ {
				var s int
				for i := 0; i < g.Width; i++ {
					for j := 0; j < g.Height; j++ {
						for q := 0; q < g.Channels; q++ {
							a.Set(r, s, f.At(i+u, j+v, q))
							s++
						}
					}
				}
				r++
			}
		}
	}
	x := blas.NewMat(m*n*k, K)
	{
		var r int
		for i := 0; i < g.Width; i++ {
			for j := 0; j < g.Height; j++ {
				for q := 0; q < g.Channels; q++ {
					for p := 0; p < h.Channels; p++ {
						x.Set(r, p, g.Filters[p].At(i, j, q))
					}
					r++
				}
			}
		}
	}
	y := blas.MatMul(1, a, x)
	{
		var r int
		for u := 0; u < h.Width; u++ {
			for v := 0; v < h.Height; v++ {
				for p := 0; p < h.Channels; p++ {
					h.Set(u, v, p, y.At(r, p))
				}
				r++
			}
		}
	}
	return h, nil
}
Example #16
0
// CorrMultiBankStrideFFT computes the strided correlation of
// a multi-channel image with a bank of multi-channel filters.
// 	h_p[u, v] = sum_q (f_q corr g_pq)[stride*u, stride*v]
func CorrMultiBankStrideFFT(f *rimg64.Multi, g *MultiBank, stride int) (*rimg64.Multi, error) {
	out := ValidSizeStride(f.Size(), g.Size(), stride)
	if out.X <= 0 || out.Y <= 0 {
		return nil, nil
	}
	// Compute strided convolution as the sum over
	// a stride x stride grid of small convolutions.
	grid := image.Pt(stride, stride)
	// But do not divide into a larger grid than the size of the filter.
	// If the filter is smaller than the stride,
	// then some pixels in the image will not affect the output.
	grid.X = min(grid.X, g.Width)
	grid.Y = min(grid.Y, g.Height)
	// Determine the size of the sub-sampled filter.
	gsub := image.Pt(ceilDiv(g.Width, grid.X), ceilDiv(g.Height, grid.Y))
	// The sub-sampled size of the image should be such that
	// the output size is attained.
	fsub := image.Pt(out.X+gsub.X-1, out.Y+gsub.Y-1)

	// Determine optimal size for FFT.
	work, _ := FFT2Size(fsub)
	// Cache FFT of each channel of image for convolving with multiple filters.
	// Re-use plan for multiple convolutions too.
	fhat := make([]*fftw.Array2, f.Channels)
	ffwd := make([]*fftw.Plan, f.Channels)
	for k := range fhat {
		fhat[k] = fftw.NewArray2(work.X, work.Y)
		ffwd[k] = fftw.NewPlan2(fhat[k], fhat[k], fftw.Forward, fftw.Estimate)
		defer ffwd[k].Destroy()
	}
	// FFT for current filter.
	curr := fftw.NewArray2(work.X, work.Y)
	gfwd := fftw.NewPlan2(curr, curr, fftw.Forward, fftw.Estimate)
	defer gfwd.Destroy()
	// Allocate one array per output channel.
	hhat := make([]*fftw.Array2, len(g.Filters))
	for k := range hhat {
		hhat[k] = fftw.NewArray2(work.X, work.Y)
	}
	// Normalization factor.
	alpha := complex(1/float64(work.X*work.Y), 0)
	// Add the convolutions over channels and strides.
	for i := 0; i < grid.X; i++ {
		for j := 0; j < grid.Y; j++ {
			// Copy each downsampled channel and take its transform.
			for p := range fhat {
				copyChannelStrideTo(fhat[p], f, p, stride, image.Pt(i, j))
				ffwd[p].Execute()
			}
			for q := range hhat {
				for p := range fhat {
					copyChannelStrideTo(curr, g.Filters[q], p, stride, image.Pt(i, j))
					gfwd.Execute()
					addMul(hhat[q], fhat[p], curr)
				}
			}
		}
	}
	// Take the inverse transform of each channel.
	h := rimg64.NewMulti(out.X, out.Y, len(g.Filters))
	for q := range hhat {
		scale(alpha, hhat[q])
		fftw.IFFT2To(hhat[q], hhat[q])
		copyRealToChannel(h, q, hhat[q])
	}
	return h, nil
}
Example #17
0
File: hog.go Project: jvlmdr/go-cv
func HOG(f *rimg64.Multi, conf Config) *rimg64.Multi {
	const eps = 0.0001
	// Leave a one-pixel border to compute derivatives.
	inside := image.Rectangle{image.ZP, f.Size()}.Inset(1)
	// Leave a half-cell border.
	half := conf.CellSize / 2
	valid := inside.Inset(half)
	// Number of whole cells inside valid region.
	cells := valid.Size().Div(conf.CellSize)
	if cells.X <= 0 || cells.Y <= 0 {
		return nil
	}
	// Remove one cell on all sides for output.
	out := cells.Sub(image.Pt(2, 2))
	// Region to iterate over.
	size := cells.Mul(conf.CellSize).Add(image.Pt(2*half, 2*half))
	vis := image.Rectangle{inside.Min, inside.Min.Add(size)}

	// Accumulate edges into cell histograms.
	hist := rimg64.NewMulti(cells.X, cells.Y, 2*conf.Angles)
	quantizer := makeQuantizer(conf.Angles)
	for a := vis.Min.X; a < vis.Max.X; a++ {
		for b := vis.Min.Y; b < vis.Max.Y; b++ {
			x, y := a-half-vis.Min.X, b-half-vis.Min.Y
			// Pick channel with strongest gradient.
			grad, v := maxGrad(f, a, b)
			v = math.Sqrt(v)
			// Snap to orientation.
			q := quantizer.quantize(grad)

			// Add to 4 histograms around pixel using bilinear interpolation.
			xp := (float64(x)+0.5)/float64(conf.CellSize) - 0.5
			yp := (float64(y)+0.5)/float64(conf.CellSize) - 0.5
			// Extract integer and fractional part.
			ixp, vx0 := modf(xp)
			iyp, vy0 := modf(yp)
			// Complement of fraction part.
			vx1 := 1 - vx0
			vy1 := 1 - vy0

			if ixp >= 0 && iyp >= 0 {
				addToMulti(hist, ixp, iyp, q, vx1*vy1*v)
			}
			if ixp+1 < cells.X && iyp >= 0 {
				addToMulti(hist, ixp+1, iyp, q, vx0*vy1*v)
			}
			if ixp >= 0 && iyp+1 < cells.Y {
				addToMulti(hist, ixp, iyp+1, q, vx1*vy0*v)
			}
			if ixp+1 < cells.X && iyp+1 < cells.Y {
				addToMulti(hist, ixp+1, iyp+1, q, vx0*vy0*v)
			}
		}
	}

	// compute energy in each block by summing over orientations
	norm := rimg64.New(cells.X, cells.Y)
	for x := 0; x < cells.X; x++ {
		for y := 0; y < cells.Y; y++ {
			for d := 0; d < conf.Angles; d++ {
				s := hist.At(x, y, d) + hist.At(x, y, d+conf.Angles)
				addTo(norm, x, y, s*s)
			}
		}
	}

	feat := rimg64.NewMulti(out.X, out.Y, conf.Channels())
	for x := 0; x < out.X; x++ {
		for y := 0; y < out.Y; y++ {
			a, b := x+1, y+1
			// Normalization factors.
			var n [4]float64
			n[0] = 1 / math.Sqrt(adjSum(norm, a, b, a+1, b+1)+eps)
			n[1] = 1 / math.Sqrt(adjSum(norm, a, b, a+1, b-1)+eps)
			n[2] = 1 / math.Sqrt(adjSum(norm, a, b, a-1, b+1)+eps)
			n[3] = 1 / math.Sqrt(adjSum(norm, a, b, a-1, b-1)+eps)
			var off int

			// Contrast-sensitive features.
			if !conf.NoContrastVar {
				for d := 0; d < 2*conf.Angles; d++ {
					h := hist.At(a, b, d)
					var sum float64
					for _, ni := range n {
						val := h * ni
						if !conf.NoClip {
							val = math.Min(val, 0.2)
						}
						sum += val
					}
					feat.Set(x, y, off+d, sum/2)
				}
				off += 2 * conf.Angles
			}

			// Contrast-insensitive features.
			if !conf.NoContrastInvar {
				for d := 0; d < conf.Angles; d++ {
					h := hist.At(a, b, d) + hist.At(a, b, conf.Angles+d)
					var sum float64
					for _, ni := range n {
						val := h * ni
						if !conf.NoClip {
							val = math.Min(val, 0.2)
						}
						sum += val
					}
					feat.Set(x, y, off+d, sum/2)
				}
				off += conf.Angles
			}

			// Texture features.
			if !conf.NoTexture {
				for i, ni := range n {
					var sum float64
					for d := 0; d < 2*conf.Angles; d++ {
						h := hist.At(a, b, d)
						val := h * ni
						if !conf.NoClip {
							val = math.Min(val, 0.2)
						}
						sum += val
					}
					feat.Set(x, y, off+i, sum/math.Sqrt(float64(2*conf.Angles)))
				}
				off += 4
			}
		}
	}
	return feat
}