What explains the huge difference in perf with the approach from 2019 mentioned at the end? For image classification, couldn't one use the residual from applyi

What explains the huge difference in perf with the approach from 2019 mentioned at the end? For image classification, couldn't one use the residual from applying a jpeg codebook learned on a training example as metric?