How to extract raw data from Tango_train.tfrecords file?

[shubhamwagh]

Hi! I am new to tfrecord file format. Is it possible if someone can share their python code snippet to extract groundtruth pointclouds and floorplans from tfrecords file? Thanks.

[shubhamwagh]

Now I am able to extract pointclouds from tfrecords. But what I see is the pointclouds have negative RGB values (for the ones I have checked). May I know if someone has faced this problem or I am doing something wrong?

from RecordReader import *
from utils import *
import numpy as np
import argparse
from open3d import *
import cv2
import random

tfrecord_path = ["./data/Tango_train.tfrecords"]


def compute3Dcentroid(points):
    centre = np.sum(points, axis =0)/len(points)
    return centre

def normalizePoints(points):
    max_value = np.max(np.max(points))
    min_value = np.min(np.min(points))

    normalized_points = 1.0*(points - min_value)/(max_value - min_value + 1e-8)
    return normalized_points

def parse_args():
    """
    Parse input arguments
    """
    parser = argparse.ArgumentParser(description='Planenet')
    parser.add_argument('--gpu', dest='gpu_id',
                        help='GPU device id to use [0]',
                        default='0', type=str)
    #task: [train, test, predict]
    parser.add_argument('--task', dest='task',
                        help='task type: [train, test, predict]',
                        default='train', type=str)
    parser.add_argument('--restore', dest='restore',
                        help='how to restore the model',
                        default=1, type=int)
    parser.add_argument('--batchSize', dest='batchSize',
                        help='batch size',
                        default=4, type=int)
    parser.add_argument('--dataset', dest='dataset',
                        help='dataset name for test/predict',
                        default='1', type=str)
    parser.add_argument('--slice', dest='slice', help='whether or not to use the slice version.',
                        action='store_true')
    parser.add_argument('--numTestingImages', dest='numTestingImages',
                        help='the number of images to test/predict',
                        default=20, type=int)
    parser.add_argument('--fineTuningCheckpoint', dest='fineTuningCheckpoint',
                        help='specify the model for fine-tuning',
                        default='checkpoint/floornet_hybrid4_branch0123_wsf', type=str)
    parser.add_argument('--suffix', dest='suffix',
                        help='add a suffix to keyname to distinguish experiments',
                        default='', type=str)
    parser.add_argument('--l2Weight', dest='l2Weight',
                        help='L2 regulation weight',
                        default=5e-4, type=float)
    parser.add_argument('--LR', dest='LR',
                        help='learning rate',
                        default=3e-5, type=float)
    parser.add_argument('--hybrid', dest='hybrid',
                        help='hybrid training',
                        default='1', type=str)
    parser.add_argument('--branches', help='active branches of the network: 0: PointNet, 1: top-down, 2: bottom-up, 3: PointNet segmentation, 4: Image Features, 5: Image Features with Joint training, 6: Additional Layers Before Pred (0, 01, 012, 0123, 01234, 1, 02*, 013)',
                        default='0123', type=str)
    #parser.add_argument('--batch_norm', help='add batch normalization to network', action='store_true')

    parser.add_argument('--cornerLossType', dest='cornerLossType',
                        help='corner loss type',
                        default='sigmoid', type=str)
    parser.add_argument('--loss',
                        help='loss type needed. [wall corner loss, door corner loss, icon corner loss, icon segmentation, room segmentation]',
                        default='01234', type=str)
    parser.add_argument('--cornerLossWeight', dest='cornerLossWeight',
                        help='corner loss weight',
                        default=10, type=float)
    parser.add_argument('--augmentation', dest='augmentation',
                        help='augmentation (wsfd)',
                        default='wsf', type=str)
    parser.add_argument('--numPoints', dest='numPoints',
                        help='number of points',
                        default=50000, type=int)
    parser.add_argument('--numInputChannels', dest='numInputChannels',
                        help='number of input channels',
                        default=3, type=int)
    parser.add_argument('--sumScale', dest='sumScale',
                        help='avoid segment sum results to be too large',
                        default=10, type=int)
    parser.add_argument('--visualizeReconstruction', dest='visualizeReconstruction',
                        help='whether to visualize flooplan reconstruction or not',
                        default=0, type=int)
    parser.add_argument('--numFinalChannels', dest='numFinalChannels', help='the number of final channels', default=256, type=int)
    parser.add_argument('--numIterations', dest='numIterations', help='the number of iterations', default=10000, type=int)
    parser.add_argument('--startIteration', dest='startIteration', help='the index of iteration to start', default=0, type=int)
    parser.add_argument('--useCache', dest='useCache',
                        help='whether to cache or not',
                        default=1, type=int)
    parser.add_argument('--debug', dest='debug',
                        help='debug index',
                        default=-1, type=int)
    parser.add_argument('--outputLayers', dest='outputLayers',
                        help='output layers',
                        default='two', type=str)
    parser.add_argument('--kernelSize', dest='kernelSize',
                        help='corner kernel size',
                        default=11, type=int)
    parser.add_argument('--iconLossWeight', dest='iconLossWeight',
                        help='icon loss weight',
                        default=1, type=float)
    parser.add_argument('--poolingTypes', dest='poolingTypes',
                        help='pooling types',
                        default='sssmm', type=str)
    parser.add_argument('--visualize', dest='visualize',
                        help='visualize during training',
                        action='store_false')
    parser.add_argument('--iconPositiveWeight', dest='iconPositiveWeight',
                        help='icon positive weight',
                        default=10, type=int)
    parser.add_argument('--prefix', dest='prefix',
                        help='prefix',
                        default='floornet', type=str)
    parser.add_argument('--drawFinal', dest='drawFinal',
                        help='draw final',
                        action='store_false')
    parser.add_argument('--separateIconLoss', dest='separateIconLoss',
                        help='separate loss for icon',
                        action='store_false')
    parser.add_argument('--evaluateImage', dest='evaluateImage',
                        help='evaluate image',
                        action='store_true')


    args = parser.parse_args()

    #args.keyname = os.path.basename(__file__).rstrip('.py')
    #args.keyname = args.keyname.replace('train_', '')
    #layers where deep supervision happens
    #addArgs(args)
    return args

args = parse_args()

dataset_train = getDatasetTrain(tfrecord_path, args.augmentation, False, args.batchSize)
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(handle, dataset_train.output_types, dataset_train.output_shapes)

input_dict, gt_dict = iterator.get_next()

iterator_train = dataset_train.make_one_shot_iterator()

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
with tf.Session(config=config) as sess:
    handle_train = sess.run(iterator_train.string_handle())
    sess.run(tf.global_variables_initializer())
    #tf.set_random_seed(1029)

    input_dict, gt_dict = sess.run([input_dict, gt_dict], feed_dict={handle: handle_train})
    points = input_dict['points']



    #icon = gt_dict['icon']
    #room = gt_dict['room']


scan_1 = points[2,:,:3]

m = np.max(scan_1, axis=0)
#index_1 = point_indices[0]

image = drawTopDownView(scan_1,256,256)

#d = getDensity(scan_1,256,256)
#i = getDensityFromIndices(index_1)

#cv2.namedWindow('image', cv2.WINDOW_NORMAL)
#cv2.imshow('image',i)



pcd = PointCloud()
pcd.points = Vector3dVector(scan_1)
pcd.colors = Vector3dVector(points[2,:,3:6])

draw_geometries([pcd])

If you see the index from 3 to 5 which are RGB values, they are all negative.

If someone has faced this problem or if @art-programmer can give some insights on this, that will be really appreciated. Thanks.