Skip to content

Latest commit

 

History

History
64 lines (53 loc) · 2.06 KB

datasetFormat.md

File metadata and controls

64 lines (53 loc) · 2.06 KB

Dataset Format for Training V^2-Net

Before training V^2-Net on your own dataset, you should add your dataset information to the datasets directory. A dataset contains a dataset splits file and several sub-dataset files. For example, we have added the ETH-UCY and SDD dataset files in the datasets folder:

datasets
|___eth.plist
|___hotel.plist
|___sdd.plist
|___univ.plist
|___zara1.plist
|___zara2.plist
|___subsets
    |___...

Dataset Splits File

It contains the dataset splits used for training and evaluation. For example, you can save the following python dict object as the MyDataset.plist (Maybe a python package like biplist is needed):

my_dataset = {
'test': ['test_subset1'],
'train': ['train_subset1', 'train_subset2', 'train_subset3'],
'val': ['val_subset1', 'val_subset2'],
}

Sub-Dataset File

You should edit and put information about all your sub-dataset that you have written into the dataset splits file into the /datasets/subsets directory. For example, you can save the following python dict object as the test_subset1.plist:

test_subset1 = {
'dataset': 'test_subset1',    # name of that sub-dataset
'dataset_dir': '....',        # root dir for your dataset csv file
'order': [1, 0],              # x-y order in your csv file
'paras': [1, 30],             # [your data fps, your video fps]
'scale': 1,                   # scale when save visualization figs
'video_path': '....',         # path for the corresponding video file 
}

Besides, all trajectories should be saved in the following true_pos_.csv format:

  • Size of the matrix is 4 x numTrajectoryPoints
  • The first row contains all the frame numbers
  • The second row contains all the pedestrian IDs
  • The third row contains all the y-coordinates
  • The fourth row contains all the x-coordinates