Model Training Dataset

DAVIS Dataset

Library currently has two preprocessing method

1. Preprocessing Method in BIRNAT

Crop, rotate, flip and etc., on image dataset before training your model and store the augmented data.

Click DAVIS_mat download augmented dataset, then modify data_root value in configs/_base_/birnat_davis.py file, make sure data_root link to your training dataset path.

Do not do any augmentation while training. Codes can be found in cacti/datasets/birnat_davis.py file.

2. Data Augmentation when Training

Download DAVIS 2017 Dataset from DAVIS website, then modify data_root value in configs/_base_/davis.py file, make sure data_root link to your training dataset path.

Do not do any augmentation before training. Crop, scale, flip and etc., while training. Codes can be found in cacti/datasets/pipelines/augmentation.py file.

Moreover, users can modify their own preprocessing function in the augmentation.py file. (e.g. random scaling)

@PIPELINES.register_module
class RandomResize:
    def __init__(self,scale=(0.8,1.2)):
        pass
    def __call__(self, imgs):
        pass

Add new preprocessing function in configs/_base_/daivs.py

#Add or delete dataaugmentation method
train_pipeline = [
    ... 
    dict(type='RandomResize',scale=(0.6,1.4))
] 

Create new cacti/datasets/davis.py file

from .builder import DATASETS

@DATASETS.register_module 
class DavisData(Dataset):
    def __init__(self,data_root,mask_path):
        pass
    def __getitem__(self,index):
        pass

Notice: function _getitem_ returns to gt as default.

Add following command in cacti/datasets/_init_.py file

from .davis import  DavisData

Create configs/_base_/davis.py file

#Add or delete dataaugmentation method
train_pipeline = [
    dict(type='Flip', direction='horizontal',flip_ratio=0.5,),
    dict(type='Resize',resize_h=256,resize_w=256)
]

gene_meas = dict(type='GenerationGrayMeas') 

train_data = dict(
    type="DavisData",
    data_root="", #dataset path 
    mask_path="test_datasets/mask/mask.mat", #mask path 
    pipeline=train_pipeline, #data augmentation 
    gene_meas = gene_meas, #generate measurement and gt 
    mask_shape = None
    #If mask_path=None, set mask shape here
)