def tests(): global times px = [] p_np = [] p_c = [] p_cl = [] iters = 15 for pixels in range(10, 480, 10): px += [pixels * pixels] print pixels rect = ((0, 0), (pixels, pixels)) p_np += [ timeit( lambda: normals.normals_numpy(depth, rect), number=iters, ) ] p_c += [ timeit( lambda: normals.normals_c(depth, rect), number=iters, ) ] p_cl += [ timeit( lambda: normals.normals_opencl(depth, rect), number=iters, ) ] times = np.vstack((px, p_np, p_c, p_cl)) times[1:, :] /= iters
def tests(): global times px = [] p_np = [] p_c = [] p_cl = [] iters=15 for pixels in range(10,480,10): px += [pixels*pixels] print pixels rect = ((0,0),(pixels,pixels)) p_np += [timeit(lambda:normals.normals_numpy(depth,rect), number=iters,)] p_c += [timeit(lambda:normals.normals_c(depth,rect), number=iters,)] p_cl += [timeit(lambda:normals.normals_opencl(depth,rect), number=iters,)] times = np.vstack((px,p_np,p_c,p_cl)) times[1:,:] /= iters
def find_plane(depth, boundpts): from visuals.camerawindow import CameraWindow global window if not 'window' in globals(): window = CameraWindow() # Build a mask of the image inside the convex points clicked u,v = uv = np.mgrid[:480,:640][::-1] mask = np.ones((480,640),bool) for (x,y),(dx,dy) in zip(boundpts, boundpts - np.roll(boundpts,1,0)): mask &= ((uv[0]-x)*dy - (uv[1]-y)*dx)<0 # Borrow the initialization from calibkinect KK = np.linalg.inv(calibkinect.projection()).astype('f') KK = np.ascontiguousarray(KK) # Find the average plane going through here global n,w n,w = normals.normals_c(depth) maskw = mask & (w>0) abc = n[maskw].mean(0) abc /= np.sqrt(np.dot(abc,abc)) a,b,c = abc x,y,z = [_[maskw].mean() for _ in calibkinect.convertOpenNI2Real_numpy(depth)] d = -(a*x+b*y+c*z) tableplane = np.array([a,b,c,d]) #tablemean = np.array([x,y,z]) # Backproject the table plane into the image using inverse transpose global tb0 tb0 = np.dot(KK.transpose(), tableplane) tb0[2] = tb0[2] # Build a matrix projecting sensor points to an system with # the origin on the table, and Y pointing up from the table # NOTE: the default orientation is offset by 45 degrees from the camera's # natural direction. v1 = np.array([a,b,c]); v1 /= np.sqrt(np.dot(v1,v1)) v0 = np.cross(v1, [-1,0,1]); v0 /= np.sqrt(np.dot(v0,v0)) v2 = np.cross(v0,v1); Ktable = np.eye(4) Ktable[:3,:3] = np.vstack((v0,v1,v2)).transpose() Ktable[:3,3] = [x,y,z] Ktable = np.linalg.inv(Ktable).astype('f') KtableKK = np.dot(Ktable, KK).astype('f') global boundptsM boundptsM = [] for (up,vp) in boundpts: # First project the image points (u,v) onto to the (imaged) tableplane dp = -(tb0[0]*up + tb0[1]*vp + tb0[3])/tb0[2] # Then project them into metric space xp, yp, zp, wp = np.dot(KtableKK, [up,vp,dp,1]) xp /= wp ; yp /= wp ; zp /= wp; boundptsM += [[xp,yp,zp]] # Use OpenGL and framebuffers to draw the table and the walls fbo = glGenFramebuffers(1) glBindFramebuffer(GL_FRAMEBUFFER, fbo); rb,rbc = glGenRenderbuffers(2) glBindRenderbuffer(GL_RENDERBUFFER, rb); glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT24, 640, 480) glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, rb); glEnable(GL_DEPTH_TEST) glClear(GL_DEPTH_BUFFER_BIT) glViewport(0, 0, 640, 480) glMatrixMode(GL_MODELVIEW) glLoadIdentity() glMatrixMode(GL_PROJECTION) glLoadIdentity() glOrtho(0,640,0,480,0,-10) glMultMatrixf(np.linalg.inv(KtableKK).transpose()) def draw(): glClear(GL_DEPTH_BUFFER_BIT|GL_COLOR_BUFFER_BIT) glEnable(GL_CULL_FACE) glBegin(GL_QUADS) for x,y,z in boundptsM: glVertex(x,y,z,1) for (x,y,z),(x_,y_,z_) in zip(boundptsM,np.roll(boundptsM,1,0)): glVertex(x ,y ,z ,1) glVertex(x_,y_,z_,1) glVertex(0,1,0,0) glVertex(0,1,0,0) glEnd() glDisable(GL_CULL_FACE) glFinish() gf = glGetIntegerv(GL_FRONT_FACE) glFrontFace(GL_CCW) draw() openglbgHi = glReadPixels(0, 0, 640, 480, GL_DEPTH_COMPONENT, GL_FLOAT).reshape(480,640); glFrontFace(GL_CW) draw() openglbgLo = glReadPixels(0, 0, 640, 480, GL_DEPTH_COMPONENT, GL_FLOAT).reshape(480,640); glFrontFace(gf) global hi,lo openglbgHi = 1000./(openglbgHi*10) openglbgLo = 1000./(openglbgLo*10) lo = np.array(openglbgLo) hi = np.array(openglbgHi) #openglbgLo[openglbgLo>=2047] = 0 #openglbgHi[np.isnan(openglbgHi)] = 0 openglbgLo[openglbgHi==openglbgLo] = 0 glBindFramebuffer(GL_FRAMEBUFFER, 0); glDeleteRenderbuffers(2, [rb,rbc]); glDeleteFramebuffers(1, [fbo]); background = np.array(depth) background[~mask] = 0 background = np.maximum(background,openglbgHi) #backgroundM = normals.project(background) openglbgLo = openglbgLo.astype(np.uint16) background = background.astype(np.uint16) background[background>=5] -= 5 #openglbgLo += 5 return dict( bgLo=openglbgLo, bgHi=background, boundpts=boundpts, boundptsM=boundptsM, KK=KK, Ktable=Ktable)
def find_plane(depth, boundpts): from visuals.camerawindow import CameraWindow global window if not 'window' in globals(): window = CameraWindow() # Build a mask of the image inside the convex points clicked u, v = uv = np.mgrid[:480, :640][::-1] mask = np.ones((480, 640), bool) for (x, y), (dx, dy) in zip(boundpts, boundpts - np.roll(boundpts, 1, 0)): mask &= ((uv[0] - x) * dy - (uv[1] - y) * dx) < 0 # Borrow the initialization from calibkinect KK = np.linalg.inv(calibkinect.projection()).astype('f') KK = np.ascontiguousarray(KK) # Find the average plane going through here global n, w n, w = normals.normals_c(depth) maskw = mask & (w > 0) abc = n[maskw].mean(0) abc /= np.sqrt(np.dot(abc, abc)) a, b, c = abc x, y, z = [ _[maskw].mean() for _ in calibkinect.convertOpenNI2Real_numpy(depth) ] d = -(a * x + b * y + c * z) tableplane = np.array([a, b, c, d]) #tablemean = np.array([x,y,z]) # Backproject the table plane into the image using inverse transpose global tb0 tb0 = np.dot(KK.transpose(), tableplane) tb0[2] = tb0[2] # Build a matrix projecting sensor points to an system with # the origin on the table, and Y pointing up from the table # NOTE: the default orientation is offset by 45 degrees from the camera's # natural direction. v1 = np.array([a, b, c]) v1 /= np.sqrt(np.dot(v1, v1)) v0 = np.cross(v1, [-1, 0, 1]) v0 /= np.sqrt(np.dot(v0, v0)) v2 = np.cross(v0, v1) Ktable = np.eye(4) Ktable[:3, :3] = np.vstack((v0, v1, v2)).transpose() Ktable[:3, 3] = [x, y, z] Ktable = np.linalg.inv(Ktable).astype('f') KtableKK = np.dot(Ktable, KK).astype('f') global boundptsM boundptsM = [] for (up, vp) in boundpts: # First project the image points (u,v) onto to the (imaged) tableplane dp = -(tb0[0] * up + tb0[1] * vp + tb0[3]) / tb0[2] # Then project them into metric space xp, yp, zp, wp = np.dot(KtableKK, [up, vp, dp, 1]) xp /= wp yp /= wp zp /= wp boundptsM += [[xp, yp, zp]] # Use OpenGL and framebuffers to draw the table and the walls fbo = glGenFramebuffers(1) glBindFramebuffer(GL_FRAMEBUFFER, fbo) rb, rbc = glGenRenderbuffers(2) glBindRenderbuffer(GL_RENDERBUFFER, rb) glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT24, 640, 480) glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, rb) glEnable(GL_DEPTH_TEST) glClear(GL_DEPTH_BUFFER_BIT) glViewport(0, 0, 640, 480) glMatrixMode(GL_MODELVIEW) glLoadIdentity() glMatrixMode(GL_PROJECTION) glLoadIdentity() glOrtho(0, 640, 0, 480, 0, -10) glMultMatrixf(np.linalg.inv(KtableKK).transpose()) def draw(): glClear(GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT) glEnable(GL_CULL_FACE) glBegin(GL_QUADS) for x, y, z in boundptsM: glVertex(x, y, z, 1) for (x, y, z), (x_, y_, z_) in zip(boundptsM, np.roll(boundptsM, 1, 0)): glVertex(x, y, z, 1) glVertex(x_, y_, z_, 1) glVertex(0, 1, 0, 0) glVertex(0, 1, 0, 0) glEnd() glDisable(GL_CULL_FACE) glFinish() gf = glGetIntegerv(GL_FRONT_FACE) glFrontFace(GL_CCW) draw() openglbgHi = glReadPixels(0, 0, 640, 480, GL_DEPTH_COMPONENT, GL_FLOAT).reshape(480, 640) glFrontFace(GL_CW) draw() openglbgLo = glReadPixels(0, 0, 640, 480, GL_DEPTH_COMPONENT, GL_FLOAT).reshape(480, 640) glFrontFace(gf) global hi, lo openglbgHi = 1000. / (openglbgHi * 10) openglbgLo = 1000. / (openglbgLo * 10) lo = np.array(openglbgLo) hi = np.array(openglbgHi) #openglbgLo[openglbgLo>=2047] = 0 #openglbgHi[np.isnan(openglbgHi)] = 0 openglbgLo[openglbgHi == openglbgLo] = 0 glBindFramebuffer(GL_FRAMEBUFFER, 0) glDeleteRenderbuffers(2, [rb, rbc]) glDeleteFramebuffers(1, [fbo]) background = np.array(depth) background[~mask] = 0 background = np.maximum(background, openglbgHi) #backgroundM = normals.project(background) openglbgLo = openglbgLo.astype(np.uint16) background = background.astype(np.uint16) background[background >= 5] -= 5 #openglbgLo += 5 return dict(bgLo=openglbgLo, bgHi=background, boundpts=boundpts, boundptsM=boundptsM, KK=KK, Ktable=Ktable)
def find_plane(depth, boundpts): from wxpy3d.camerawindow import CameraWindow global window if not 'window' in globals(): window = CameraWindow() # Build a mask of the image inside the convex points clicked mask = make_mask(boundpts) # Borrow the initialization from calibkinect KK = np.linalg.inv(calibkinect.projection()).astype('f') KK = np.ascontiguousarray(KK) # Find the average plane going through here global n,w n,w = normals.normals_c(depth) maskw = mask & (w>0) abc = n[maskw].mean(0) abc /= np.sqrt(np.dot(abc,abc)) a,b,c = abc x,y,z = [_[maskw].mean() for _ in calibkinect.convertOpenNI2Real_numpy(depth)] d = -(a*x+b*y+c*z) tableplane = np.array([a,b,c,d]) #tablemean = np.array([x,y,z]) # Backproject the table plane into the image using inverse transpose global tb0 tb0 = np.dot(KK.T, tableplane) tb0[2] = tb0[2] # Build a matrix projecting sensor points to an system with # the origin on the table, and Y pointing up from the table # NOTE: the default orientation is offset by 45 degrees from the camera's # natural direction. v1 = np.array([a,b,c]); v1 /= np.sqrt(np.dot(v1,v1)) v0 = np.cross(v1, [-1,0,1]); v0 /= np.sqrt(np.dot(v0,v0)) v2 = np.cross(v0,v1); Ktable = np.eye(4) Ktable[:3,:3] = np.vstack((v0,v1,v2)).T Ktable[:3,3] = [x,y,z] Ktable = np.linalg.inv(Ktable).astype('f') KtableKK = np.dot(Ktable, KK).astype('f') global boundptsM boundptsM = [] for (up,vp) in boundpts: # First project the image points (u,v) onto to the (imaged) tableplane dp = -(tb0[0]*up + tb0[1]*vp + tb0[3])/tb0[2] # Then project them into metric space xp, yp, zp, wp = np.dot(KtableKK, [up,vp,dp,1]) xp /= wp ; yp /= wp ; zp /= wp; boundptsM += [[xp,yp,zp]] # Use OpenGL and framebuffers to draw the table and the walls fbo = glGenFramebuffers(1) glBindFramebuffer(GL_FRAMEBUFFER, fbo); rb,rbc = glGenRenderbuffers(2) glBindRenderbuffer(GL_RENDERBUFFER, rb); glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT24, 640, 480) glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, rb); glEnable(GL_DEPTH_TEST) glClear(GL_DEPTH_BUFFER_BIT) glViewport(0, 0, 640, 480) glMatrixMode(GL_MODELVIEW) glLoadIdentity() glMatrixMode(GL_PROJECTION) glLoadIdentity() glOrtho(0,640,0,480,0,-10) glMultMatrixf(np.linalg.inv(KtableKK).T) def draw(): glClear(GL_DEPTH_BUFFER_BIT|GL_COLOR_BUFFER_BIT) glEnable(GL_CULL_FACE) glBegin(GL_QUADS) for x,y,z in boundptsM: glVertex(x,y,z,1) for (x,y,z),(x_,y_,z_) in zip(boundptsM,np.roll(boundptsM,1,0)): glVertex(x ,y ,z ,1) glVertex(x_,y_,z_,1) glVertex(0,1,0,0) glVertex(0,1,0,0) glEnd() glDisable(GL_CULL_FACE) glFinish() # Rendering the outside faces gives us the near walls gf = glGetIntegerv(GL_FRONT_FACE) glFrontFace(GL_CCW) draw() openglbgHi = glReadPixels(0, 0, 640, 480, GL_DEPTH_COMPONENT, GL_FLOAT).reshape(480,640) # Rendering the interior faces gives us the table plane and the far walls glFrontFace(GL_CW) draw() openglbgLo = glReadPixels(0, 0, 640, 480, GL_DEPTH_COMPONENT, GL_FLOAT).reshape(480,640) glFrontFace(gf) # We need to invert the depth measurements to obtain kinect-style range # images. # initially, openglbgHi/Lo are in units of 1/m. # afterwards, they are in units of mm. global hi,lo openglbgHi = 1000./(openglbgHi*10) openglbgLo = 1000./(openglbgLo*10) lo = np.array(openglbgLo) hi = np.array(openglbgHi) openglbgLo[openglbgHi==openglbgLo] = 0 glBindFramebuffer(GL_FRAMEBUFFER, 0); glDeleteRenderbuffers(2, [rb,rbc]); glDeleteFramebuffers(1, [fbo]); # Some of the observed image points are nearer than the fitted plane. # We want fewer false positives in this case, so take the maximum # of either the fitted plane or the observed depth background = np.array(depth) background[~mask] = 0 background = np.maximum(background,openglbgHi) openglbgLo = openglbgLo.astype(np.uint16) background = background.astype(np.uint16) background[background>=5] -= 5 # Reduce false positives even more. return dict( bgLo=openglbgLo, bgHi=background, boundpts=boundpts, boundptsM=boundptsM, KK=KK, Ktable=Ktable)
def find_plane(depth, boundpts): from wxpy3d.camerawindow import CameraWindow global window if not 'window' in globals(): window = CameraWindow() # Build a mask of the image inside the convex points clicked mask = make_mask(boundpts) # Borrow the initialization from calibkinect KK = np.linalg.inv(calibkinect.projection()).astype('f') KK = np.ascontiguousarray(KK) # Find the average plane going through here global n, w n, w = normals.normals_c(depth) maskw = mask & (w > 0) abc = n[maskw].mean(0) abc /= np.sqrt(np.dot(abc, abc)) a, b, c = abc x, y, z = [ _[maskw].mean() for _ in calibkinect.convertOpenNI2Real_numpy(depth) ] d = -(a * x + b * y + c * z) tableplane = np.array([a, b, c, d]) #tablemean = np.array([x,y,z]) # Backproject the table plane into the image using inverse transpose global tb0 tb0 = np.dot(KK.T, tableplane) tb0[2] = tb0[2] # Build a matrix projecting sensor points to an system with # the origin on the table, and Y pointing up from the table # NOTE: the default orientation is offset by 45 degrees from the camera's # natural direction. v1 = np.array([a, b, c]) v1 /= np.sqrt(np.dot(v1, v1)) v0 = np.cross(v1, [-1, 0, 1]) v0 /= np.sqrt(np.dot(v0, v0)) v2 = np.cross(v0, v1) Ktable = np.eye(4) Ktable[:3, :3] = np.vstack((v0, v1, v2)).T Ktable[:3, 3] = [x, y, z] Ktable = np.linalg.inv(Ktable).astype('f') KtableKK = np.dot(Ktable, KK).astype('f') global boundptsM boundptsM = [] for (up, vp) in boundpts: # First project the image points (u,v) onto to the (imaged) tableplane dp = -(tb0[0] * up + tb0[1] * vp + tb0[3]) / tb0[2] # Then project them into metric space xp, yp, zp, wp = np.dot(KtableKK, [up, vp, dp, 1]) xp /= wp yp /= wp zp /= wp boundptsM += [[xp, yp, zp]] # Use OpenGL and framebuffers to draw the table and the walls fbo = glGenFramebuffers(1) glBindFramebuffer(GL_FRAMEBUFFER, fbo) rb, rbc = glGenRenderbuffers(2) glBindRenderbuffer(GL_RENDERBUFFER, rb) glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT24, 640, 480) glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, rb) glEnable(GL_DEPTH_TEST) glClear(GL_DEPTH_BUFFER_BIT) glViewport(0, 0, 640, 480) glMatrixMode(GL_MODELVIEW) glLoadIdentity() glMatrixMode(GL_PROJECTION) glLoadIdentity() glOrtho(0, 640, 0, 480, 0, -10) glMultMatrixf(np.linalg.inv(KtableKK).T) def draw(): glClear(GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT) glEnable(GL_CULL_FACE) glBegin(GL_QUADS) for x, y, z in boundptsM: glVertex(x, y, z, 1) for (x, y, z), (x_, y_, z_) in zip(boundptsM, np.roll(boundptsM, 1, 0)): glVertex(x, y, z, 1) glVertex(x_, y_, z_, 1) glVertex(0, 1, 0, 0) glVertex(0, 1, 0, 0) glEnd() glDisable(GL_CULL_FACE) glFinish() # Rendering the outside faces gives us the near walls gf = glGetIntegerv(GL_FRONT_FACE) glFrontFace(GL_CCW) draw() openglbgHi = glReadPixels(0, 0, 640, 480, GL_DEPTH_COMPONENT, GL_FLOAT).reshape(480, 640) # Rendering the interior faces gives us the table plane and the far walls glFrontFace(GL_CW) draw() openglbgLo = glReadPixels(0, 0, 640, 480, GL_DEPTH_COMPONENT, GL_FLOAT).reshape(480, 640) glFrontFace(gf) # We need to invert the depth measurements to obtain kinect-style range # images. # initially, openglbgHi/Lo are in units of 1/m. # afterwards, they are in units of mm. global hi, lo openglbgHi = 1000. / (openglbgHi * 10) openglbgLo = 1000. / (openglbgLo * 10) lo = np.array(openglbgLo) hi = np.array(openglbgHi) openglbgLo[openglbgHi == openglbgLo] = 0 glBindFramebuffer(GL_FRAMEBUFFER, 0) glDeleteRenderbuffers(2, [rb, rbc]) glDeleteFramebuffers(1, [fbo]) # Some of the observed image points are nearer than the fitted plane. # We want fewer false positives in this case, so take the maximum # of either the fitted plane or the observed depth background = np.array(depth) background[~mask] = 0 background = np.maximum(background, openglbgHi) openglbgLo = openglbgLo.astype(np.uint16) background = background.astype(np.uint16) background[background >= 5] -= 5 # Reduce false positives even more. return dict(bgLo=openglbgLo, bgHi=background, boundpts=boundpts, boundptsM=boundptsM, KK=KK, Ktable=Ktable)