v1std.motion_dog and corresponding test

Author: rcoreilly

examples/motion_dog/motion_dog.go

@@ -23,6 +23,7 @@ import (
 	"github.com/emer/emergent/v2/edge"
 	"github.com/emer/v1vision/dog"
 	"github.com/emer/v1vision/motion"
+	"github.com/emer/v1vision/v1std"
 	"github.com/emer/v1vision/v1vision"
 )
 
@@ -94,6 +95,9 @@ type Vis struct { //types:add
 	// FullField integrated output
 	FullField tensor.Float32 `display:"no-inline"`
 
+	// MotionDoG is a self-contained version of motion dog filtering.
+	MotionDoG v1std.MotionDoG
+
 	fastIdx, starIdx int
 
 	tabView *core.Tabs
@@ -113,11 +117,10 @@ func (vi *Vis) Defaults() {
 	sz := 12 // V1mF16 typically = 12, no border
 	spc := 4
 	vi.DoG.SetSize(sz, spc)
-	// note: first arg is border -- we are relying on Geom
-	// to set border to .5 * filter size
-	// any further border sizes on same image need to add Geom.FiltRt!
 	vi.Geom.Set(math32.Vec2i(0, 0), math32.Vec2i(spc, spc), math32.Vec2i(sz, sz))
 	vi.Geom.SetImageSize(vi.ImageSize)
+	vi.MotionDoG.Defaults()
+	// vi.MotionDoG.GPU = false
 }
 
 func (vi *Vis) Config() {
@@ -139,6 +142,7 @@ func (vi *Vis) Config() {
 	if vi.GPU {
 		vi.V1.GPUInit()
 	}
+	vi.MotionDoG.Config(vi.ImageSize)
 }
 
 // RenderFrames renders the frames
@@ -147,17 +151,12 @@ func (vi *Vis) RenderFrames() { //types:add
 	vi.Motion.NormInteg = 0
 	vi.Pos = vi.Start
 	for i := range vi.NFrames {
-		_ = i
 		vi.RenderFrame()
 		vi.Pos = vi.Pos.Add(vi.Velocity)
 		vi.Filter()
 		if vi.tabView != nil {
-			// fmt.Println(i)
 			vi.tabView.AsyncLock()
 			vi.tabView.Update()
-			// vi.fastView.Update()
-			// vi.slowView.Update()
-			// vi.imgView.Update()
 			vi.tabView.AsyncUnlock()
 			fmt.Printf("%d\tL: %7.4g\tR: %7.4g\tB: %7.4g\tT: %7.4g\tN: %7.4g\n", i, vi.FullField.Value1D(0), vi.FullField.Value1D(1), vi.FullField.Value1D(2), vi.FullField.Value1D(3), vi.Motion.NormInteg)
 			time.Sleep(vi.FrameDelay)
@@ -183,6 +182,9 @@ func (vi *Vis) RenderFrame() {
 // Filter runs the filters on current image.
 func (vi *Vis) Filter() error { //types:add
 	v1vision.UseGPU = vi.GPU
+	vi.V1.SetAsCurrent()
+	// note: if switching between different gpu, then zero values get copied
+	// up to GPU on SetAsCurrent, so need ValuesVar to synchronize.
 	vi.V1.Run(v1vision.ScalarsVar, v1vision.ValuesVar, v1vision.ImagesVar)
 	// vi.V1.Run(v1vision.ScalarsVar) // minimal fastest case
 	vi.Motion.FullFieldInteg(vi.V1.Scalars, &vi.FullField)
@@ -203,6 +205,8 @@ func (vi *Vis) Filter() error { //types:add
 	vi.Star.SetShapeSizes(int(vi.Geom.Out.Y-1), int(vi.Geom.Out.X-1), 2, 4)
 	tensor.CopyFromLargerShape(&vi.Star, star)
 
+	vi.MotionDoG.RunTensor(vi.ImageTsr)
+
 	return nil
 }
 
@@ -221,7 +225,7 @@ func (vi *Vis) ConfigGUI() *core.Body {
 		s.Range.FixMax = false
 	})
 
-	b := core.NewBody("lgn_dog").SetTitle("LGN DoG Filtering")
+	b := core.NewBody("lgn_dog").SetTitle("Motion DoG Filtering")
 	sp := core.NewSplits(b)
 	core.NewForm(sp).SetStruct(vi)
 	tb := core.NewTabs(sp)

motion/enumgen.go

@@ -14,6 +14,17 @@ import (
 
 //go:generate core generate -add-types -gosl
 
+// Directions are the motion directions, in feature order,
+// as represented in the Star and FullField outputs.
+type Directions int32 //enums:enum
+
+const (
+	Left Directions = iota
+	Right
+	Down
+	Up
+)
+
 // Params has the motion parameters for retinal starburst amacrine
 // cells (SAC) that compute centrifugal motion flow from each point.
 type Params struct {

motion/motion.go

@@ -51,19 +51,23 @@ type DoGColor struct {
 func (vi *DoGColor) Defaults() {
 	vi.GPU = true
 	vi.DoG.Defaults()
-	sz := 12  // V1mF16 typically = 12, no border -- defaults
-	spc := 16 // note: not 4; broader blob tuning
-	vi.DoG.Spacing = spc
-	vi.DoG.Size = sz
 	vi.DoG.Gain = 8 // color channels are weaker than grey
 	vi.DoG.OnGain = 1
 	vi.DoG.SetSameSigma(0.5) // no spatial component, just pure contrast
-	vi.Geom.Set(math32.Vec2i(0, 0), math32.Vec2i(spc, spc), math32.Vec2i(sz, sz))
+	vi.SetSize(12, 16)       // V1mF16 typically = 12, no border
 	vi.KWTA.Defaults()
 	vi.KWTA.Layer.On.SetBool(false) // non-spatial, mainly for differentiation within pools
 	vi.KWTA.Pool.Gi = 1.2
 }
 
+// SetSize sets the V1sGabor filter size and geom spacing to given values.
+// Default is 12, 16, for a medium-sized filter.
+func (vi *DoGColor) SetSize(sz, spc int) {
+	vi.DoG.Spacing = spc
+	vi.DoG.Size = sz
+	vi.Geom.Set(math32.Vec2i(0, 0), math32.Vec2i(spc, spc), math32.Vec2i(sz, sz))
+}
+
 // Config configures the filtering pipeline with all the current parameters.
 // imageSize is the _content_ size of input image that is passed to Run
 // as an RGB Tensor (per [V1Vision.Images] standard format),
@@ -80,7 +84,7 @@ func (vi *DoGColor) Config(imageSize image.Point) {
 	lmsRG := vi.V1.NewImage(vi.Geom.In.V())
 	lmsBY := vi.V1.NewImage(vi.Geom.In.V())
 
-	vi.V1.NewWrapImage(img, 3, wrap, int(vi.Geom.FilterRt.X), &vi.Geom)
+	vi.V1.NewWrapImage(img, 3, wrap, int(vi.Geom.Border.X), &vi.Geom)
 	vi.V1.NewLMSComponents(wrap, lmsRG, lmsBY, vi.DoG.Gain, &vi.Geom)
 
 	out := vi.V1.NewValues(int(vi.Geom.Out.Y), int(vi.Geom.Out.X), 2)

motion/typegen.go

@@ -40,8 +40,12 @@ type DoGGrey struct {
 func (vi *DoGGrey) Defaults() {
 	vi.GPU = true
 	vi.DoG.Defaults()
-	spc := 4
-	sz := 12
+	vi.SetSize(12, 4)
+}
+
+// SetSize sets the V1sGabor filter size and geom spacing to given values.
+// Default is 12, 4, for a medium-sized filter.
+func (vi *DoGGrey) SetSize(sz, spc int) {
 	vi.DoG.Spacing = spc
 	vi.DoG.Size = sz
 	vi.Geom.Set(math32.Vec2i(0, 0), math32.Vec2i(spc, spc), math32.Vec2i(sz, sz))
@@ -59,7 +63,7 @@ func (vi *DoGGrey) Config(imageSize image.Point) {
 	img := vi.V1.NewImage(vi.Geom.In.V())
 	wrap := vi.V1.NewImage(vi.Geom.In.V())
 
-	vi.V1.NewWrapImage(img, 0, wrap, int(vi.Geom.FilterRt.X), &vi.Geom)
+	vi.V1.NewWrapImage(img, 0, wrap, int(vi.Geom.Border.X), &vi.Geom)
 	_, out := vi.V1.NewDoG(wrap, 0, &vi.DoG, &vi.Geom)
 	vi.V1.NewLogValues(out, out, 1, 1.0, &vi.Geom)
 	vi.V1.NewNormDiv(v1vision.MaxScalar, out, out, 1, &vi.Geom)

v1std/dog_color.go

@@ -0,0 +1,145 @@
+// Copyright (c) 2025, The Emergent Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package v1std
+
+import (
+	"image"
+
+	"cogentcore.org/core/math32"
+	"cogentcore.org/lab/tensor"
+	"github.com/emer/v1vision/dog"
+	"github.com/emer/v1vision/motion"
+	"github.com/emer/v1vision/v1vision"
+)
+
+// MotionDoG computes starburst-amacrine style motion processing and
+// resulting summary full-field motion values, on greyscale
+// difference-of-gaussian (DoG) filtering.
+// Call Defaults and then set any custom params, then call Config.
+// Results are in Output tensor after Run().
+type MotionDoG struct {
+	// GPU means use the GPU by default (does GPU initialization) in Config.
+	// To change what is actually used at the moment of running,
+	// set [v1vision.UseGPU].
+	GPU bool
+
+	// LGN DoG filter parameters.
+	DoG dog.Filter
+
+	// Motion filter parameters.
+	Motion motion.Params
+
+	// Geom is geometry of input, output.
+	Geom v1vision.Geom `edit:"-"`
+
+	// FullField has the integrated FullField output (1D).
+	// Use [motion.Directions] for 1D indexes (is 2x2 for [L,R][D,U]).
+	FullField tensor.Float32 `display:"no-inline"`
+
+	// GetStar retrieves the star values. Otherwise, just the full-field.
+	GetStar bool
+
+	// V1 is the V1Vision filter processing system.
+	V1 v1vision.V1Vision `display:"no-inline"`
+
+	// Star has the star values, if GetStar is true,
+	// pointing to Values in V1.
+	// [Y, X, Polarity, 4], where Polarity is DoG polarity, and 4 is for
+	// Left, Right, Down, Up.
+	Star *tensor.Float32 `display:"no-inline"`
+}
+
+func (vi *MotionDoG) Defaults() {
+	vi.GPU = true
+	vi.DoG.Defaults()
+	vi.Motion.Defaults()
+	vi.SetSize(12, 4)
+}
+
+// SetSize sets the V1sGabor filter size and geom spacing to given values.
+// Default is 12, 4, for a medium-sized filter.
+func (vi *MotionDoG) SetSize(sz, spc int) {
+	vi.DoG.Spacing = spc
+	vi.DoG.Size = sz
+	vi.Geom.Set(math32.Vec2i(0, 0), math32.Vec2i(spc, spc), math32.Vec2i(sz, sz))
+}
+
+// Config configures the filtering pipeline with all the current parameters.
+// imageSize is the _content_ size of input image that is passed to Run
+// as an RGB Tensor (per [V1Vision.Images] standard format),
+// (i.e., exclusive of the additional border around the image = [Image.Size]).
+// The resulting Geom.Border field can be passed to [Image] methods.
+func (vi *MotionDoG) Config(imageSize image.Point) {
+	vi.Geom.SetImageSize(imageSize)
+	vi.FullField.SetShapeSizes(2, 2)
+
+	fn := 1 // number of filters in DoG
+
+	vi.V1.Init()
+	img := vi.V1.NewImage(vi.Geom.In.V())
+	wrap := vi.V1.NewImage(vi.Geom.In.V())
+
+	vi.V1.NewWrapImage(img, 0, wrap, int(vi.Geom.Border.X), &vi.Geom)
+	_, out := vi.V1.NewDoG(wrap, 0, &vi.DoG, &vi.Geom)
+	vi.V1.NewLogValues(out, out, fn, 1.0, &vi.Geom)
+	vi.V1.NewNormDiv(v1vision.MaxScalar, out, out, fn, &vi.Geom)
+
+	vi.Motion.DoGSumScalarIndex = vi.V1.NewAggScalar(v1vision.SumScalar, out, fn, &vi.Geom)
+	fastIdx := vi.V1.NewMotionIntegrate(out, fn, vi.Motion.FastTau, vi.Motion.SlowTau, &vi.Geom)
+	starIdx := vi.V1.NewMotionStar(fastIdx, fn, vi.Motion.Gain, &vi.Geom)
+	vi.Motion.FFScalarIndex = vi.V1.NewMotionFullField(starIdx, fn, &vi.Geom)
+
+	if vi.GetStar {
+		out4 := vi.V1.NewValues4D(int(vi.Geom.Out.Y), int(vi.Geom.Out.X), 2, 4)
+		vi.Star.SetShapeSizes(int(vi.Geom.Out.Y), int(vi.Geom.Out.X), 2, 4)
+		vi.V1.NewTo4D(starIdx, out4, 2, 4, 0, &vi.Geom)
+	}
+
+	vi.V1.SetAsCurrent()
+	if vi.GPU {
+		vi.V1.GPUInit()
+	}
+}
+
+// RunImage runs the configured filtering pipeline
+// on given Image, using given [Image] handler.
+func (vi *MotionDoG) RunImage(im *Image, img image.Image) {
+	im.SetImageGrey(&vi.V1, img, int(vi.Geom.Border.X))
+	vi.Run()
+}
+
+// Run runs the configured filtering pipeline
+// on given Image tensor.
+func (vi *MotionDoG) RunTensor(tsr *tensor.Float32) {
+	itsr := vi.V1.Images.SubSpace(0).(*tensor.Float32)
+	itsr.CopyFrom(tsr)
+	vi.Run()
+}
+
+// Run runs the configured filtering pipeline.
+// image in vi.V1.Images[0] must already have been set.
+func (vi *MotionDoG) Run() {
+	v1vision.UseGPU = vi.GPU
+	vi.V1.SetAsCurrent()
+
+	vals := []v1vision.GPUVars{v1vision.ScalarsVar, v1vision.ValuesVar}
+	if vi.GetStar {
+		vals = append(vals, v1vision.Values4DVar)
+	}
+	vi.V1.Run(vals...)
+	vi.Motion.FullFieldInteg(vi.V1.Scalars, &vi.FullField)
+	if vi.GetStar {
+		vi.Star = vi.V1.Values4D.SubSpace(0).(*tensor.Float32)
+	}
+}
+
+// Init resets all motion integration values to 0.
+func (vi *MotionDoG) Init() {
+	vi.V1.SetAsCurrent()
+	tensor.SetAllFloat64(vi.V1.Values, 0)
+	tensor.SetAllFloat64(vi.V1.Scalars, 0)
+	tensor.SetAllFloat64(&vi.FullField, 0)
+	vi.V1.ToGPUInfra()
+}

v1std/dog_grey.go

@@ -94,7 +94,7 @@ func (vi *V1cColor) Config(imageSize image.Point) {
 	wrap := vi.V1.NewImage(vi.V1sGeom.In.V())
 	lms := vi.V1.NewImage(vi.V1sGeom.In.V())
 
-	vi.fadeOpIdx = vi.V1.NewFadeImage(img, 3, wrap, int(vi.V1sGeom.FilterRt.X), .5, .5, .5, &vi.V1sGeom)
+	vi.fadeOpIdx = vi.V1.NewFadeImage(img, 3, wrap, int(vi.V1sGeom.Border.X), .5, .5, .5, &vi.V1sGeom)
 	vi.V1.NewLMSOpponents(wrap, lms, vi.ColorGain, &vi.V1sGeom)
 
 	nang := vi.V1sGabor.NAngles
@@ -161,7 +161,7 @@ func (vi *V1cColor) RunImage(im *Image, img image.Image) {
 	vi.V1.SetAsCurrent()
 	v1vision.UseGPU = vi.GPU
 	im.SetImageRGB(&vi.V1, img, int(vi.V1sGeom.Border.X))
-	r, g, b := v1vision.EdgeAvg(im.Tsr, int(vi.V1sGeom.FilterRt.X))
+	r, g, b := v1vision.EdgeAvg(im.Tsr, int(vi.V1sGeom.Border.X))
 	vi.V1.SetFadeRGB(vi.fadeOpIdx, r, g, b)
 	vi.V1.Run(v1vision.Values4DVar)
 	vi.Output = vi.V1.Values4D.SubSpace(0).(*tensor.Float32)