how to test myself dataset?

We will update the documentation afterward. In Vis4D, a data pipeline has several components: a dataset class, a data pipe, a data loader, and a data connector.

Take the COCO detection dataset as an example. First, you need to implement your custom dataset class like COCO here, which is responsible for loading data and annotations with pre-defined common keys.

Then, through the example here, you can compose your dataset along with the desired data transform with our implemented data pipe and data loader to work with our trainer.

Finally, we rely on the data connector to propagate the data/output between the data loader, model, evaluator, and visualizer. As the example here shows, these connectors connect data between the dataset and model. The dictionary key is the args for the model input, while the value is the key for the data dictionary after data augmentation.

More specifically, the following data connector will map the K.images from the data loader output to images args in the model forward.

CONN_BBOX_2D_TEST = {
    "images": K.images,
    "input_hw": K.input_hw,
    "original_hw": K.original_hw,
}

SysCV / vis4d

how to test myself dataset? #137

Describe the Feature

Motivation

Pitch

Alternatives

Additional context