Code to create a PyTorch data loader for datasets in Supervisely format. To train Faster-RCNN as well as Keypoint-RCNN.

As of July 7th 2020 this code has been tested using:

Supervisely datasets are stored in the following folder structure:

/dataset_name/ann/img01.png.json
/dataset_name/ann/img02.png.json
/dataset_name/ann/...
/dataset_name/img/img01.png
/dataset_name/img/img02.png
/dataset_name/img/...
/dataset_name/meta.json
/dataset_name/obj_class_to_machine_color.json

Keep this structure (only the two folders 'ann' and 'img' are needed).

For bounding box detection with Faster-RCNN use:

from superviselyboxes import SuperviselyBoxes
DATA_ROOT = '/path/to/data/'

# Initialize dataset
train_data = SuperviselyBoxes(root = DATA_ROOT, dataset_name = 'train', transform=None)

# Create data loader
train_data_loader = torch.utils.data.DataLoader(
    train_data, batch_size=4, shuffle=True, num_workers=4,
    collate_fn=collate_fn
)

# Fetch next batch
imgs, targets = next(iter(train_data_loader))

Note: collate_fn is taken from torch vision/references/detection/utils.py
Just use this code:

def collate_fn(batch):
    return tuple(zip(*batch))

In FasterRCNN-HowTo-Example.ipynb you find a full tutorial on how to use the dataloader to train a Faster-RCNN in PyTorch.
It also includes code for visualization of the image and it's annotations.

• Support loading mask annotations from polygon as well as bitmap geometryType
• Making the loader work with arbitrary transforms