Agfdhyk

I'm trying to override the __getitem__ method in the PyTorch Dataset abstract class for an object detection task where the dataset consists of multiple images, each of which has multiple bounding boxes.

https://pytorch.org/tutorials/beginner/data_loading_tutorial.html

The bounding boxes are defined by 5 parameters: the upper-left coordinate of the rectangle, the lower right coordinate of the rectangle, and the class label annotation for the object detected inside the bounding box.

So: x1, y1, x2, y2, class_label

The PyTorch __getitem__ method is supposed to return a single (image, set_of_bounding_box_and_annotations) when called with an index of integer type. So __getitem__(17) would return the 17th image with the set of all bounding-boxes and labels.

The format of the data is a dictionary of lists, each of which consists of a dictionary.

Ex:

my_dict =

{img_1.png: [{x1: 0, y1: 0, x2: 10, y2: 20, label: 'dog'},  {x1: 30, y1: 40, x2: 50, y2: 60, label: 'cat'}, ...],

{img_2.png: [{x1: 84, y1: 27, x2: 95, y2: 43, label: 'bird'}, {x1: 91, y1: 91, x2: 102, y2: 110, label: 'alligator'}, ...],

...

}

So PyTorch requires the __getitem__ method to return an image,box-set sample when passing it the index of the data.

Problem is a dictionary has no inherent order in Python, so you can't just call my_dict[0]. Seems like an OrderedDict would work here, but we want the value of the key-value pair to be the set/list of boundingbox/label dictionaries.

How to get the data into an indexable format that the __getitem__ method can return a sample (image, bbox_and_labels)?