Probabilistic Embeddings Revisited

TL;DR

This paper critically analyzes probabilistic embedding methods in face verification, proposing improvements that set new benchmarks, and explores confidence prediction's relation to data quality and error estimation.

Contribution

It provides the first comprehensive analysis of probabilistic methods in verification, introduces a simple extension achieving state-of-the-art results, and establishes a benchmark for confidence prediction evaluation.

Findings

Proposed a simple extension that improves probabilistic verification performance.

Confidence correlates with data quality but poorly predicts error probability.

Established a new benchmark for confidence prediction in verification tasks.

Abstract

In recent years, deep metric learning and its probabilistic extensions claimed state-of-the-art results in the face verification task. Despite improvements in face verification, probabilistic methods received little attention in the research community and practical applications. In this paper, we, for the first time, perform an in-depth analysis of known probabilistic methods in verification and retrieval tasks. We study different design choices and propose a simple extension, achieving new state-of-the-art results among probabilistic methods. Finally, we study confidence prediction and show that it correlates with data quality, but contains little information about prediction error probability. We thus provide a new confidence evaluation benchmark and establish a baseline for future confidence prediction research. PyTorch implementation is publicly released.