Does Progress On Object Recognition Benchmarks Improve Real-World Generalization?