When an image has no predictions, it would be nice if utilities.format_geometry natively processed this into the output. Right now we have to do something like

predictions = m.predict_step(images, batch_index)
for pred in predictions:
    geometry = determine_geometry(pred)
    if len(pred[geometry]) == 0:
            y_pred = {}
            y_pred["y"] = torch.zeros(4)
            y_pred["labels"] = torch.zeros(1)
            y_pred["scores"] = torch.zeros(1)
    else:
            format_geometry(pred, geom_type=geometry)

Desired functionality

1. a new single function 'format_prediction' that looks at a prediction dictionary, uses determine_geometry, format_geometry, and fills in the edge case of empty images.
2. Minimal tests that shows this behavior for points and boxes, @jveitchmichaelis we don't have polygon prediction yet, right?