Beispiel #1
0
def runVisualSfm(self, imageCollectionId, sceneId, cleanup=True, history=None):
  from voxel_globe.meta import models
  from voxel_globe.order.visualsfm.models import Order

  from os import environ as env
  from os.path import join as path_join
  import os
  import shutil
  
  from .tools import writeNvm, writeGcpFile, generateMatchPoints, runSparse,\
                     readNvm
  
  import voxel_globe.tools
  from voxel_globe.tools.wget import download as wget
  from voxel_globe.tools.camera import get_kto
  import voxel_globe.tools.enu as enu
  import numpy

  import boxm2_adaptor
  import boxm2_scene_adaptor
  from voxel_globe.tools.xml_dict import load_xml
  
  from django.contrib.gis.geos import Point
  from voxel_globe.tools.image import convert_image

  from distutils.spawn import find_executable

  from glob import glob
  
  self.update_state(state='INITIALIZE', meta={'stage':0})

  #Make main temp dir and cd into it
  with voxel_globe.tools.task_dir('visualsfm', cd=True) as processing_dir:

    #Because visualsfm is so... bad, I have to copy it locally so I can
    #configure it
    visualsfm_exe = os.path.join(processing_dir, 
        os.path.basename(os.environ['VIP_VISUALSFM_EXE']))
    shutil.copy(find_executable(os.environ['VIP_VISUALSFM_EXE']), 
                visualsfm_exe)
    with open(os.path.join(processing_dir, 'nv.ini'), 'w') as fid:
      fid.write('param_search_multiple_models 0\n')
      fid.write('param_use_siftgpu 2\n')

    matchFilename = path_join(processing_dir, 'match.nvm');
    sparce_filename = path_join(processing_dir, 'sparse.nvm');
    #This can NOT be changed in version 0.5.25  
    gcpFilename = matchFilename + '.gcp'
    logger.debug('Task %s is processing in %s' % (self.request.id, 
                                                  processing_dir))

    image_collection = models.ImageCollection.objects.get(
        id=imageCollectionId).history(history);
    imageList = image_collection.images.all();

    #A Little bit of database logging
    oid = Order(processingDir=processing_dir, imageCollection=image_collection)

###    if 1:
###    try: #Not fully integrated yet
###      sift_gpu = siftgpu.SiftGPU()
###    except:
###      pass

    localImageList = [];
    for x in range(len(imageList)):
      #Download the image locally
      image = imageList[x].history(history);
      self.update_state(state='INITIALIZE', meta={'stage':'image fetch', 'i':x,
                                                  'total':len(imageList)})
      imageName = image.originalImageUrl;
      extension = os.path.splitext(imageName)[1].lower()
      localName = path_join(processing_dir, 'frame_%05d%s' % (x+1, extension));
      wget(imageName, localName, secret=True)
  
      #Convert the image if necessary    
      if extension not in ['.jpg', '.jpeg', '.pgm', '.ppm']:
        self.update_state(state='INITIALIZE', 
            meta={'stage':'image convert', 'i':x, 'total':len(imageList)})
        #Add code here to converty to jpg for visual sfm
        if extension in ['.png']:#'not implemented':
          from PIL import Image
          image_temp = Image.open(localName)
          if len(image_temp.mode) > 1: #Stupid visual sfm is picky :(
            new_local_name = os.path.splitext(localName)[0] + '.ppm';
          else:
            new_local_name = os.path.splitext(localName)[0] + '.pgm';

          new_local_name = os.path.splitext(localName)[0] + '.jpg';

          ###ingest.convert_image(localName, new_local_name, 'PNM')
          convert_image(localName, new_local_name, 'JPEG', 
                        options=('QUALITY=100',))
          os.remove(localName)

          localName = new_local_name;

        else:
          raise Exception('Unsupported file type');
        
      imageInfo = {'localName':localName, 'index':x}
  
      try:
        [K, T, llh] = get_kto(image, history=history);
        imageInfo['K_intrinsics'] = K;
        imageInfo['transformation'] = T;
        imageInfo['enu_origin'] = llh;
      except:
        pass
  
      localImageList.append(imageInfo);
###      if 1:
###      try: #not fully integrated yet
###        sift_gpu.create_sift(localName, os.path.splitext(localName)[0]+'.sift')
###      except:
###        pass

  #  filenames = list(imageList.values_list('imageUrl'))
  #  logger.info('The image list 0is %s' % filenames)

    self.update_state(state='PROCESSING', 
                      meta={'stage':'generate match points', 
                            'processing_dir':processing_dir,
                            'total':len(imageList)})
    generateMatchPoints(map(lambda x:x['localName'], localImageList),
                        matchFilename, logger=logger, executable=visualsfm_exe)

  #   cameras = [];
  #   for image in imageList:
  #     if 1:
  #     #try:
  #       [K, T, llh] = get_kto(image);
  #       cameras.append({'image':image.id, 'K':K, 'tranformation':
  #                       T, 'origin':llh})
  #     #except:
  #       pass  
  
  #  origin = numpy.median(origin, axis=0)
  #  origin = [-92.215197, 37.648858, 268.599]
    scene = models.Scene.objects.get(id=sceneId).history(history)
    origin = list(scene.origin)

    if scene.geolocated:
      self.update_state(state='PROCESSING', 
                        meta={'stage':'writing gcp points'})

      #find the middle origin, and make it THE origin
      data = []#.name .llh_xyz
      for imageInfo in localImageList:
        try:
          r = imageInfo['transformation'][0:3, 0:3]
          t = imageInfo['transformation'][0:3, 3:]
          enu_point = -r.transpose().dot(t);
    
          if not numpy.array_equal(imageInfo['enu_origin'], origin):
            ecef = enu.enu2xyz(refLong=imageInfo['enu_origin'][0],
                               refLat=imageInfo['enu_origin'][1],
                               refH=imageInfo['enu_origin'][2],
                               #e=imageInfo['transformation'][0, 3],
                               #n=imageInfo['transformation'][1, 3],
                               #u=imageInfo['transformation'][2, 3])
                               e=enu_point[0],
                               n=enu_point[1],
                               u=enu_point[2])
            enu_point = enu.xyz2enu(refLong=origin[0], 
                                    refLat=origin[1], 
                                    refH=origin[2],
                                    X=ecef[0],
                                    Y=ecef[1],
                                    Z=ecef[2])
    #      else:
    #        enu_point = imageInfo['transformation'][0:3, 3];
          
          dataBit = {'filename':imageInfo['localName'], 'xyz':enu_point}
          data.append(dataBit);
          
          #Make this a separate ingest process, making CAMERAS linked to the 
          #images
          #data = arducopter.loadAdjTaggedMetadata(
          #    r'd:\visualsfm\2014-03-20 13-22-44_adj_tagged_images.txt');
          #Make this read the cameras from the DB instead
          writeGcpFile(data, gcpFilename)

        except: #some images may have no camera 
          pass
    
    oid.lvcsOrigin = str(origin)
    oid.save()
 
    self.update_state(state='PROCESSING', meta={'stage':'sparse SFM'})
    runSparse(matchFilename, sparce_filename, gcp=scene.geolocated, 
              shared=True, logger=logger, executable=visualsfm_exe)
  
    self.update_state(state='FINALIZE', 
                      meta={'stage':'loading resulting cameras'})

    #prevent bundle2scene from getting confused and crashing
    sift_data = os.path.join(processing_dir, 'sift_data')
    os.mkdir(sift_data)
    for filename in glob(os.path.join(processing_dir, '*.mat')) +\
                    glob(os.path.join(processing_dir, '*.sift')):
      shutil.move(filename, sift_data)

    if scene.geolocated:
      #Create a uscene.xml for the geolocated case. All I want out of this is
      #the bounding box and gsd calculation.
      boxm2_adaptor.bundle2scene(sparce_filename, processing_dir, isalign=False,
                                 out_dir="")

      cams = readNvm(path_join(processing_dir, 'sparse.nvm'))
      #cams.sort(key=lambda x:x.name)
      #Since the file names are frame_00001, etc... and you KNOW this order is
      #identical to localImageList, with some missing
      for cam in cams:
        frameName = cam.name; #frame_00001, etc....
        imageInfo = filter(lambda x: x['localName'].endswith(frameName),
                           localImageList)[0]
        #I have to use endswith instead of == because visual sfm APPARENTLY 
        #decides to take some liberty and make absolute paths relative
        image = imageList[imageInfo['index']].history(history)
    
        (k,r,t) = cam.krt(width=image.imageWidth, height=image.imageHeight);
        logger.info('Origin is %s' % str(origin))
        llh_xyz = enu.enu2llh(lon_origin=origin[0], 
                              lat_origin=origin[1], 
                              h_origin=origin[2], 
                              east=cam.translation_xyz[0], 
                              north=cam.translation_xyz[1], 
                              up=cam.translation_xyz[2])
            
        grcs = models.GeoreferenceCoordinateSystem.create(
                        name='%s 0' % image.name,
                        xUnit='d', yUnit='d', zUnit='m',
                        location='SRID=4326;POINT(%0.15f %0.15f %0.15f)' 
                                  % (origin[0], origin[1], origin[2]),
                        service_id = self.request.id)
        grcs.save()
        cs = models.CartesianCoordinateSystem.create(
                        name='%s 1' % (image.name),
                        service_id = self.request.id,
                        xUnit='m', yUnit='m', zUnit='m');
        cs.save()

        transform = models.CartesianTransform.create(
                             name='%s 1_0' % (image.name),
                             service_id = self.request.id,
                             rodriguezX=Point(*r[0,:]),
                             rodriguezY=Point(*r[1,:]),
                             rodriguezZ=Point(*r[2,:]),
                             translation=Point(t[0][0], t[1][0], t[2][0]),
                             coordinateSystem_from_id=grcs.id,
                             coordinateSystem_to_id=cs.id)
        transform.save()
        
        camera = image.camera;
        try:
          camera.update(service_id = self.request.id,
                        focalLengthU=k[0,0],   focalLengthV=k[1,1],
                        principalPointU=k[0,2], principalPointV=k[1,2],
                        coordinateSystem=cs);
        except:
          camera = models.Camera.create(name=image.name,
                        service_id = self.request.id,
                        focalLengthU=k[0,0],   focalLengthV=k[1,1],
                        principalPointU=k[0,2], principalPointV=k[1,2],
                        coordinateSystem=cs);
          camera.save();
          image.update(camera = camera);
    
      logger.info(str(cams[0]))
    else:
      from vsi.tools.natural_sort import natural_sorted 
      from glob import glob
      
      from vsi.io.krt import Krt
      from voxel_globe.tools.camera import save_krt
      
      boxm2_adaptor.bundle2scene(sparce_filename, processing_dir, isalign=True,
                                 out_dir=processing_dir)
      #While the output dir is used for the b2s folders, uscene.xml is cwd
      #They are both set to processing_dir, so everything works out well
      aligned_cams = glob(os.path.join(processing_dir, 'cams_krt', '*'))
      #sort them naturally in case there are more then 99,999 files
      aligned_cams = natural_sorted(aligned_cams) 
      if len(aligned_cams) != len(imageList):
        #Create a new image collection
        new_image_collection = models.ImageCollection.create(
            name="SFM Result Subset (%s)" % image_collection.name, 
            service_id = self.request.id);
#        for image in image_collection.images.all():
#          new_image_collection.images.add(image)
        new_image_collection.save();

        frames_keep = set(map(lambda x:
            int(os.path.splitext(x.split('_')[-2])[0])-1, aligned_cams))

        for frame_index in frames_keep:
          new_image_collection.images.add(imageList[frame_index])

#        frames_remove = set(xrange(len(imageList))) - frames_keep 
#
#        for remove_index in list(frames_remove):
#          #The frame number refers to the nth image in the image collection,
#          #so frame_00100.tif is the 100th image, starting the index at one
#          #See local_name above
#          
#          #remove the images sfm threw away 
#          new_image_collection.remove(imageList[remove_index])
        image_collection = new_image_collection
        frames_keep = list(frames_keep)
      else:
        frames_keep = xrange(len(aligned_cams))
      
      #---Update the camera models in the database.---
      for camera_index, frame_index in enumerate(frames_keep):
        krt = Krt.load(aligned_cams[camera_index])
        image = imageList[frame_index].history(history)
        save_krt(self.request.id, image, krt.k, krt.r, krt.t, [0,0,0], 
                 srid=4326)

      #---Update scene information important for the no-metadata case ---

    scene_filename = os.path.join(processing_dir, 'model', 'uscene.xml')
    boxm_scene = boxm2_scene_adaptor.boxm2_scene_adaptor(scene_filename)

    scene.bbox_min = 'POINT(%0.15f %0.15f %0.15f)' % boxm_scene.bbox[0]
    scene.bbox_max = 'POINT(%0.15f %0.15f %0.15f)' % boxm_scene.bbox[1]

    #This is not a complete or good function really... but it will get me the
    #information I need.
    scene_dict = load_xml(scene_filename)
    block = scene_dict['block']

    scene.default_voxel_size='POINT(%f %f %f)' % \
        (float(block.at['dim_x']), float(block.at['dim_y']),
         float(block.at['dim_z']))
    scene.save()

  return oid.id;
Beispiel #2
0
def runVisualSfm(self, imageCollectionId, sceneId, cleanup=True, history=None):
    from voxel_globe.meta import models
    from voxel_globe.order.visualsfm.models import Order

    from os import environ as env
    from os.path import join as path_join
    import os
    import shutil

    from .tools import writeNvm, writeGcpFile, generateMatchPoints, runSparse,\
                       readNvm

    import voxel_globe.tools
    from voxel_globe.tools.wget import download as wget
    from voxel_globe.tools.camera import get_kto
    import voxel_globe.tools.enu as enu
    import numpy

    import boxm2_adaptor
    import boxm2_scene_adaptor
    from voxel_globe.tools.xml_dict import load_xml

    from django.contrib.gis.geos import Point
    from voxel_globe.tools.image import convert_image

    from distutils.spawn import find_executable

    from glob import glob

    self.update_state(state='INITIALIZE', meta={'stage': 0})

    #Make main temp dir and cd into it
    with voxel_globe.tools.task_dir('visualsfm', cd=True) as processing_dir:

        #Because visualsfm is so... bad, I have to copy it locally so I can
        #configure it
        visualsfm_exe = os.path.join(
            processing_dir, os.path.basename(os.environ['VIP_VISUALSFM_EXE']))
        shutil.copy(find_executable(os.environ['VIP_VISUALSFM_EXE']),
                    visualsfm_exe)
        with open(os.path.join(processing_dir, 'nv.ini'), 'w') as fid:
            fid.write('param_search_multiple_models 0\n')
            fid.write('param_use_siftgpu 2\n')

        matchFilename = path_join(processing_dir, 'match.nvm')
        sparce_filename = path_join(processing_dir, 'sparse.nvm')
        #This can NOT be changed in version 0.5.25
        gcpFilename = matchFilename + '.gcp'
        logger.debug('Task %s is processing in %s' %
                     (self.request.id, processing_dir))

        image_collection = models.ImageCollection.objects.get(
            id=imageCollectionId).history(history)
        imageList = image_collection.images.all()

        #A Little bit of database logging
        oid = Order(processingDir=processing_dir,
                    imageCollection=image_collection)

        ###    if 1:
        ###    try: #Not fully integrated yet
        ###      sift_gpu = siftgpu.SiftGPU()
        ###    except:
        ###      pass

        localImageList = []
        for x in range(len(imageList)):
            #Download the image locally
            image = imageList[x].history(history)
            self.update_state(state='INITIALIZE',
                              meta={
                                  'stage': 'image fetch',
                                  'i': x,
                                  'total': len(imageList)
                              })
            imageName = image.originalImageUrl
            extension = os.path.splitext(imageName)[1].lower()
            localName = path_join(processing_dir,
                                  'frame_%05d%s' % (x + 1, extension))
            wget(imageName, localName, secret=True)

            #Convert the image if necessary
            if extension not in ['.jpg', '.jpeg', '.pgm', '.ppm']:
                self.update_state(state='INITIALIZE',
                                  meta={
                                      'stage': 'image convert',
                                      'i': x,
                                      'total': len(imageList)
                                  })
                #Add code here to converty to jpg for visual sfm
                if extension in ['.png']:  #'not implemented':
                    from PIL import Image
                    image_temp = Image.open(localName)
                    if len(image_temp.mode
                           ) > 1:  #Stupid visual sfm is picky :(
                        new_local_name = os.path.splitext(
                            localName)[0] + '.ppm'
                    else:
                        new_local_name = os.path.splitext(
                            localName)[0] + '.pgm'

                    new_local_name = os.path.splitext(localName)[0] + '.jpg'

                    ###ingest.convert_image(localName, new_local_name, 'PNM')
                    convert_image(localName,
                                  new_local_name,
                                  'JPEG',
                                  options=('QUALITY=100', ))
                    os.remove(localName)

                    localName = new_local_name

                else:
                    raise Exception('Unsupported file type')

            imageInfo = {'localName': localName, 'index': x}

            try:
                [K, T, llh] = get_kto(image, history=history)
                imageInfo['K_intrinsics'] = K
                imageInfo['transformation'] = T
                imageInfo['enu_origin'] = llh
            except:
                pass

            localImageList.append(imageInfo)
###      if 1:
###      try: #not fully integrated yet
###        sift_gpu.create_sift(localName, os.path.splitext(localName)[0]+'.sift')
###      except:
###        pass

#  filenames = list(imageList.values_list('imageUrl'))
#  logger.info('The image list 0is %s' % filenames)

        self.update_state(state='PROCESSING',
                          meta={
                              'stage': 'generate match points',
                              'processing_dir': processing_dir,
                              'total': len(imageList)
                          })
        generateMatchPoints(map(lambda x: x['localName'], localImageList),
                            matchFilename,
                            logger=logger,
                            executable=visualsfm_exe)

        #   cameras = [];
        #   for image in imageList:
        #     if 1:
        #     #try:
        #       [K, T, llh] = get_kto(image);
        #       cameras.append({'image':image.id, 'K':K, 'tranformation':
        #                       T, 'origin':llh})
        #     #except:
        #       pass

        #  origin = numpy.median(origin, axis=0)
        #  origin = [-92.215197, 37.648858, 268.599]
        scene = models.Scene.objects.get(id=sceneId).history(history)
        origin = list(scene.origin)

        if scene.geolocated:
            self.update_state(state='PROCESSING',
                              meta={'stage': 'writing gcp points'})

            #find the middle origin, and make it THE origin
            data = []  #.name .llh_xyz
            for imageInfo in localImageList:
                try:
                    r = imageInfo['transformation'][0:3, 0:3]
                    t = imageInfo['transformation'][0:3, 3:]
                    enu_point = -r.transpose().dot(t)

                    if not numpy.array_equal(imageInfo['enu_origin'], origin):
                        ecef = enu.enu2xyz(
                            refLong=imageInfo['enu_origin'][0],
                            refLat=imageInfo['enu_origin'][1],
                            refH=imageInfo['enu_origin'][2],
                            #e=imageInfo['transformation'][0, 3],
                            #n=imageInfo['transformation'][1, 3],
                            #u=imageInfo['transformation'][2, 3])
                            e=enu_point[0],
                            n=enu_point[1],
                            u=enu_point[2])
                        enu_point = enu.xyz2enu(refLong=origin[0],
                                                refLat=origin[1],
                                                refH=origin[2],
                                                X=ecef[0],
                                                Y=ecef[1],
                                                Z=ecef[2])
        #      else:
        #        enu_point = imageInfo['transformation'][0:3, 3];

                    dataBit = {
                        'filename': imageInfo['localName'],
                        'xyz': enu_point
                    }
                    data.append(dataBit)

                    #Make this a separate ingest process, making CAMERAS linked to the
                    #images
                    #data = arducopter.loadAdjTaggedMetadata(
                    #    r'd:\visualsfm\2014-03-20 13-22-44_adj_tagged_images.txt');
                    #Make this read the cameras from the DB instead
                    writeGcpFile(data, gcpFilename)

                except:  #some images may have no camera
                    pass

        oid.lvcsOrigin = str(origin)
        oid.save()

        self.update_state(state='PROCESSING', meta={'stage': 'sparse SFM'})
        runSparse(matchFilename,
                  sparce_filename,
                  gcp=scene.geolocated,
                  shared=True,
                  logger=logger,
                  executable=visualsfm_exe)

        self.update_state(state='FINALIZE',
                          meta={'stage': 'loading resulting cameras'})

        #prevent bundle2scene from getting confused and crashing
        sift_data = os.path.join(processing_dir, 'sift_data')
        os.mkdir(sift_data)
        for filename in glob(os.path.join(processing_dir, '*.mat')) +\
                        glob(os.path.join(processing_dir, '*.sift')):
            shutil.move(filename, sift_data)

        if scene.geolocated:
            #Create a uscene.xml for the geolocated case. All I want out of this is
            #the bounding box and gsd calculation.
            boxm2_adaptor.bundle2scene(sparce_filename,
                                       processing_dir,
                                       isalign=False,
                                       out_dir="")

            cams = readNvm(path_join(processing_dir, 'sparse.nvm'))
            #cams.sort(key=lambda x:x.name)
            #Since the file names are frame_00001, etc... and you KNOW this order is
            #identical to localImageList, with some missing
            for cam in cams:
                frameName = cam.name
                #frame_00001, etc....
                imageInfo = filter(
                    lambda x: x['localName'].endswith(frameName),
                    localImageList)[0]
                #I have to use endswith instead of == because visual sfm APPARENTLY
                #decides to take some liberty and make absolute paths relative
                image = imageList[imageInfo['index']].history(history)

                (k, r, t) = cam.krt(width=image.imageWidth,
                                    height=image.imageHeight)
                logger.info('Origin is %s' % str(origin))
                llh_xyz = enu.enu2llh(lon_origin=origin[0],
                                      lat_origin=origin[1],
                                      h_origin=origin[2],
                                      east=cam.translation_xyz[0],
                                      north=cam.translation_xyz[1],
                                      up=cam.translation_xyz[2])

                grcs = models.GeoreferenceCoordinateSystem.create(
                    name='%s 0' % image.name,
                    xUnit='d',
                    yUnit='d',
                    zUnit='m',
                    location='SRID=4326;POINT(%0.15f %0.15f %0.15f)' %
                    (origin[0], origin[1], origin[2]),
                    service_id=self.request.id)
                grcs.save()
                cs = models.CartesianCoordinateSystem.create(
                    name='%s 1' % (image.name),
                    service_id=self.request.id,
                    xUnit='m',
                    yUnit='m',
                    zUnit='m')
                cs.save()

                transform = models.CartesianTransform.create(
                    name='%s 1_0' % (image.name),
                    service_id=self.request.id,
                    rodriguezX=Point(*r[0, :]),
                    rodriguezY=Point(*r[1, :]),
                    rodriguezZ=Point(*r[2, :]),
                    translation=Point(t[0][0], t[1][0], t[2][0]),
                    coordinateSystem_from_id=grcs.id,
                    coordinateSystem_to_id=cs.id)
                transform.save()

                camera = image.camera
                try:
                    camera.update(service_id=self.request.id,
                                  focalLengthU=k[0, 0],
                                  focalLengthV=k[1, 1],
                                  principalPointU=k[0, 2],
                                  principalPointV=k[1, 2],
                                  coordinateSystem=cs)
                except:
                    camera = models.Camera.create(name=image.name,
                                                  service_id=self.request.id,
                                                  focalLengthU=k[0, 0],
                                                  focalLengthV=k[1, 1],
                                                  principalPointU=k[0, 2],
                                                  principalPointV=k[1, 2],
                                                  coordinateSystem=cs)
                    camera.save()
                    image.update(camera=camera)

            logger.info(str(cams[0]))
        else:
            from vsi.tools.natural_sort import natural_sorted
            from glob import glob

            from vsi.io.krt import Krt
            from voxel_globe.tools.camera import save_krt

            boxm2_adaptor.bundle2scene(sparce_filename,
                                       processing_dir,
                                       isalign=True,
                                       out_dir=processing_dir)
            #While the output dir is used for the b2s folders, uscene.xml is cwd
            #They are both set to processing_dir, so everything works out well
            aligned_cams = glob(os.path.join(processing_dir, 'cams_krt', '*'))
            #sort them naturally in case there are more then 99,999 files
            aligned_cams = natural_sorted(aligned_cams)
            if len(aligned_cams) != len(imageList):
                #Create a new image collection
                new_image_collection = models.ImageCollection.create(
                    name="SFM Result Subset (%s)" % image_collection.name,
                    service_id=self.request.id)
                #        for image in image_collection.images.all():
                #          new_image_collection.images.add(image)
                new_image_collection.save()

                frames_keep = set(
                    map(
                        lambda x: int(os.path.splitext(x.split('_')[-2])[0]) -
                        1, aligned_cams))

                for frame_index in frames_keep:
                    new_image_collection.images.add(imageList[frame_index])


#        frames_remove = set(xrange(len(imageList))) - frames_keep
#
#        for remove_index in list(frames_remove):
#          #The frame number refers to the nth image in the image collection,
#          #so frame_00100.tif is the 100th image, starting the index at one
#          #See local_name above
#
#          #remove the images sfm threw away
#          new_image_collection.remove(imageList[remove_index])
                image_collection = new_image_collection
                frames_keep = list(frames_keep)
            else:
                frames_keep = xrange(len(aligned_cams))

            #---Update the camera models in the database.---
            for camera_index, frame_index in enumerate(frames_keep):
                krt = Krt.load(aligned_cams[camera_index])
                image = imageList[frame_index].history(history)
                save_krt(self.request.id,
                         image,
                         krt.k,
                         krt.r,
                         krt.t, [0, 0, 0],
                         srid=4326)

            #---Update scene information important for the no-metadata case ---

        scene_filename = os.path.join(processing_dir, 'model', 'uscene.xml')
        boxm_scene = boxm2_scene_adaptor.boxm2_scene_adaptor(scene_filename)

        scene.bbox_min = 'POINT(%0.15f %0.15f %0.15f)' % boxm_scene.bbox[0]
        scene.bbox_max = 'POINT(%0.15f %0.15f %0.15f)' % boxm_scene.bbox[1]

        #This is not a complete or good function really... but it will get me the
        #information I need.
        scene_dict = load_xml(scene_filename)
        block = scene_dict['block']

        scene.default_voxel_size='POINT(%f %f %f)' % \
            (float(block.at['dim_x']), float(block.at['dim_y']),
             float(block.at['dim_z']))
        scene.save()

    return oid.id
Beispiel #3
0
def runVisualSfm(self, imageSetId, sceneId, cleanup=True):
  from voxel_globe.meta import models

  from os import environ as env
  from os.path import join as path_join
  import os
  import shutil
  import time

  from django.contrib.gis.geos import Point
  
  from .tools import writeNvm, writeGcpFile, generateMatchPoints, runSparse,\
                     readNvm
  
  import voxel_globe.tools
  from voxel_globe.tools.camera import get_kto, save_krt
  import voxel_globe.tools.enu as enu
  import numpy

  import boxm2_adaptor
  import boxm2_scene_adaptor
  from voxel_globe.tools.xml_dict import load_xml
  
  from django.contrib.gis.geos import Point
  from voxel_globe.tools.image import convert_image

  from distutils.spawn import find_executable

  from vsi.iglob import glob as glob

  self.update_state(state='INITIALIZE', meta={'stage':0})

  #Make main temp dir and cd into it
  with voxel_globe.tools.task_dir('visualsfm', cd=True) as processing_dir:

    #Because visualsfm is so... bad, I need to copy it locally so I can
    #configure it
    visualsfm_exe = os.path.join(processing_dir, 'visualsfm')
    shutil.copy(find_executable('VisualSFM'), visualsfm_exe)
    with open(os.path.join(processing_dir, 'nv.ini'), 'w') as fid:
      fid.write('param_search_multiple_models 0\n')
      fid.write('param_use_siftgpu 2\n')

    matchFilename = path_join(processing_dir, 'match.nvm')
    sparce_filename = path_join(processing_dir, 'sparse.nvm')
    #This can NOT be changed in version 0.5.25  
    gcpFilename = matchFilename + '.gcp'
    logger.debug('Task %s is processing in %s' % (self.request.id, 
                                                  processing_dir))

    image_set = models.ImageSet.objects.get(
        id=imageSetId)
    imageList = image_set.images.all()

###    if 1:
###    try: #Not fully integrated yet
###      sift_gpu = siftgpu.SiftGPU()
###    except:
###      pass

    localImageList = []
    for x in range(len(imageList)):
      #Download the image locally
      image = imageList[x]
      self.update_state(state='INITIALIZE', meta={'stage':'image fetch', 'i':x,
                                                  'total':len(imageList)})
      imageName = image.filename_path
      extension = os.path.splitext(imageName)[1].lower()
      localName = path_join(processing_dir, 'frame_%05d%s' % (x+1, extension))
      #lncp(imageName, localName)
      #Stupid VisualSFM dereferences symlinks, breaking this
      shutil.copyfile(imageName, localName)
  
      #Convert the image if necessary    
      if extension not in ['.jpg', '.jpeg', '.pgm', '.ppm']:
        self.update_state(state='INITIALIZE', 
            meta={'stage':'image convert', 'i':x, 'total':len(imageList)})
        #Add code here to converty to jpg for visual sfm
        if extension in ['.png']:#'not implemented':
          from PIL import Image
          image_temp = Image.open(localName)
          # if len(image_temp.mode) > 1: #Stupid visual sfm is picky :(
          #   new_local_name = os.path.splitext(localName)[0] + '.ppm'
          # else:
          #   new_local_name = os.path.splitext(localName)[0] + '.pgm'

          new_local_name = os.path.splitext(localName)[0] + '.jpg'

          ###ingest.convert_image(localName, new_local_name, 'PNM')
          convert_image(localName, new_local_name, 'JPEG', 
                        options=('QUALITY=100',))
          os.remove(localName)

          localName = new_local_name

        else:
          raise Exception('Unsupported file type')
        
      imageInfo = {'localName':localName, 'index':x}
  
      try:
        [K, T, llh] = get_kto(image)
        imageInfo['K_intrinsics'] = K
        imageInfo['transformation'] = T
        imageInfo['enu_origin'] = llh
      except:
        pass
  
      localImageList.append(imageInfo)
###      if 1:
###      try: #not fully integrated yet
###        sift_gpu.create_sift(localName, os.path.splitext(localName)[0]+'.sift')
###      except:
###        pass

  #  filenames = list(imageList.values_list('image_url'))
  #  logger.info('The image list 0is %s' % filenames)

    self.update_state(state='PROCESSING', 
                      meta={'stage':'generate match points', 
                            'processing_dir':processing_dir,
                            'total':len(imageList)})
    pid = generateMatchPoints(map(lambda x:x['localName'], localImageList),
                              matchFilename, logger=logger,
                              executable=visualsfm_exe)
    
    old_mat=None
    old_sift=None

    #TODO: Replace with inotify to monitor directory
    while pid.poll() is None:
      mat = len(glob(os.path.join(processing_dir, '*.mat'), False))
      sift = len(glob(os.path.join(processing_dir, '*.sift'), False))
      if mat  != old_mat or \
         sift != old_sift:
        old_mat=mat
        old_sift=sift
        self.update_state(state='PROCESSING', 
                          meta={'stage':'generate match points', 
                                'processing_dir':processing_dir,
                                'sift':sift,
                                'mat':mat,
                                'total':len(imageList)})
      time.sleep(5)

  #   cameras = []
  #   for image in imageList:
  #     if 1:
  #     #try:
  #       [K, T, llh] = get_kto(image)
  #       cameras.append({'image':image.id, 'K':K, 'tranformation':
  #                       T, 'origin':llh})
  #     #except:
  #       pass  
  
  #  origin = numpy.median(origin, axis=0)
  #  origin = [-92.215197, 37.648858, 268.599]
    scene = models.Scene.objects.get(id=sceneId)
    origin = list(scene.origin)

    if scene.geolocated:
      self.update_state(state='PROCESSING', 
                        meta={'stage':'writing gcp points'})

      #find the middle origin, and make it THE origin
      data = []#.name .llh_xyz
      for imageInfo in localImageList:
        try:
          r = imageInfo['transformation'][0:3, 0:3]
          t = imageInfo['transformation'][0:3, 3:]
          enu_point = -r.transpose().dot(t)
    
          if not numpy.array_equal(imageInfo['enu_origin'], origin):
            ecef = enu.enu2xyz(refLong=imageInfo['enu_origin'][0],
                               refLat=imageInfo['enu_origin'][1],
                               refH=imageInfo['enu_origin'][2],
                               #e=imageInfo['transformation'][0, 3],
                               #n=imageInfo['transformation'][1, 3],
                               #u=imageInfo['transformation'][2, 3])
                               e=enu_point[0],
                               n=enu_point[1],
                               u=enu_point[2])
            enu_point = enu.xyz2enu(refLong=origin[0], 
                                    refLat=origin[1], 
                                    refH=origin[2],
                                    X=ecef[0],
                                    Y=ecef[1],
                                    Z=ecef[2])
    #      else:
    #        enu_point = imageInfo['transformation'][0:3, 3]
          
          dataBit = {'filename':imageInfo['localName'], 'xyz':enu_point}
          data.append(dataBit)
          
          #Make this a separate ingest process, making CAMERAS linked to the 
          #images
          #data = arducopter.loadAdjTaggedMetadata(
          #    r'd:\visualsfm\2014-03-20 13-22-44_adj_tagged_images.txt')
          #Make this read the cameras from the DB instead
          writeGcpFile(data, gcpFilename)

        except: #some images may have no camera 
          pass
    
    self.update_state(state='PROCESSING', meta={'stage':'sparse SFM'})
    pid = runSparse(matchFilename, sparce_filename, gcp=scene.geolocated, 
                    shared=True, logger=logger, executable=visualsfm_exe)
    pid.wait()
  
    self.update_state(state='FINALIZE', 
                      meta={'stage':'loading resulting cameras'})

    #prevent bundle2scene from getting confused and crashing
    sift_data = os.path.join(processing_dir, 'sift_data')
    os.mkdir(sift_data)
    for filename in glob(os.path.join(processing_dir, '*.mat'), False) +\
                    glob(os.path.join(processing_dir, '*.sift'), False):
      shutil.move(filename, sift_data)

    if scene.geolocated:
      #Create a uscene.xml for the geolocated case. All I want out of this is
      #the bounding box and gsd calculation.
      boxm2_adaptor.bundle2scene(sparce_filename, processing_dir, isalign=False,
                                 out_dir="")

      cams = readNvm(path_join(processing_dir, 'sparse.nvm'))
      #cams.sort(key=lambda x:x.name)
      #Since the file names are frame_00001, etc... and you KNOW this order is
      #identical to localImageList, with some missing

      camera_set = models.CameraSet(name="Visual SFM Geo %s" % image_set.name,
                                    service_id = self.request.id,
                                    images_id = imageSetId)
      camera_set.save()

      for cam in cams:
        frameName = cam.name #frame_00001, etc....
        imageInfo = filter(lambda x: x['localName'].endswith(frameName),
                           localImageList)[0]
        #I have to use endswith instead of == because visual sfm APPARENTLY 
        #decides to take some liberty and make absolute paths relative
        image = imageList[imageInfo['index']]

        (k,r,t) = cam.krt(width=image.image_width, height=image.image_height)
        t = t.flatten()
        camera = save_krt(self.request.id, image, k, r, t, origin, srid=4326)
        camera_set.cameras.add(camera)
    else:
      from vsi.tools.natural_sort import natural_sorted 
      
      from vsi.io.krt import Krt
      
      boxm2_adaptor.bundle2scene(sparce_filename, processing_dir, isalign=True,
                                 out_dir=processing_dir)
      #While the output dir is used for the b2s folders, uscene.xml is cwd
      #They are both set to processing_dir, so everything works out well
      aligned_cams = glob(os.path.join(processing_dir, 'cams_krt', '*'))
      #sort them naturally in case there are more then 99,999 files
      aligned_cams = natural_sorted(aligned_cams) 
      if len(aligned_cams) != len(imageList):
        #Create a new image set
        new_image_set = models.ImageSet(
            name="SFM Result Subset (%s)" % image_set.name, 
            service_id = self.request.id)
#        for image in image_set.images.all():
#          new_image_set.images.add(image)
        new_image_set.save()

        frames_keep = set(map(lambda x:
            int(os.path.splitext(x.split('_')[-2])[0])-1, aligned_cams))

        for frame_index in frames_keep:
          new_image_set.images.add(imageList[frame_index])

#        frames_remove = set(xrange(len(imageList))) - frames_keep 
#
#        for remove_index in list(frames_remove):
#          #The frame number refers to the nth image in the image set,
#          #so frame_00100.tif is the 100th image, starting the index at one
#          #See local_name above
#          
#          #remove the images sfm threw away 
#          new_image_set.remove(imageList[remove_index])
        image_set = new_image_set
        frames_keep = list(frames_keep)
      else:
        frames_keep = xrange(len(aligned_cams))

      camera_set = models.CameraSet(name="Visual SFM %s" % image_set.name,
                                    service_id = self.request.id,
                                    images_id = imageSetId)
      camera_set.save()

      #---Update the camera models in the database.---
      for camera_index, frame_index in enumerate(frames_keep):
        krt = Krt.load(aligned_cams[camera_index])
        image = imageList[frame_index]
        camera = save_krt(self.request.id, image, krt.k, krt.r, krt.t, [0,0,0], 
                          srid=4326)
        camera_set.cameras.add(camera)

      #---Update scene information important for the no-metadata case ---

    scene_filename = os.path.join(processing_dir, 'model', 'uscene.xml')
    boxm_scene = boxm2_scene_adaptor.boxm2_scene_adaptor(scene_filename)

    scene.bbox_min = Point(*boxm_scene.bbox[0])
    scene.bbox_max = Point(*boxm_scene.bbox[1])

    #This is not a complete or good function really... but it will get me the
    #information I need.
    scene_dict = load_xml(scene_filename)
    block = scene_dict['block']

    scene.default_voxel_size=Point(float(block.at['dim_x']),
                                   float(block.at['dim_y']),
                                   float(block.at['dim_z']))
    scene.save()