- [ ] Check if there's a performance lift when ensembling FaceAPI, FaceNet, InsightFace, etc. - [ ] What should be the weights in the ensemble? Can they be learned? - [ ] Also investigate test-time augmentation