ODIN doesn't have an evaluation metric

[mich-431]

Currently when training ODIN the only metric that we have to evaluate is loss which is kind of hard to put in context. I'm not sure what this would look like because there are a number of things that we can evaluate against. Perhaps IOU and whether the class is correct. Another option is to make an opencv demo to test qualitatively.