Added dataset that has IRL images and a unit test to report metrics. Also adds support for missing labels for ground truth data. If an object is missing some labels, it'll only count that target toward the box precision/recall but not count the classification for that label.
The conclusion is that our shit sucks but it could probably be worse.
Summary
Added dataset that has IRL images and a unit test to report metrics. Also adds support for missing labels for ground truth data. If an object is missing some labels, it'll only count that target toward the box precision/recall but not count the classification for that label.
The conclusion is that our shit sucks but it could probably be worse.