Core Concepts
Proposing a Robust Dual Embedding method (RDE) to address noisy correspondence in Text-to-Image Person Re-identification, achieving state-of-the-art results.
Abstract
The content introduces the problem of noisy correspondence in TIReID and proposes the RDE method to mitigate its impact. It consists of Confident Consensus Division (CCD) and Triplet Alignment Loss (TAL) components. Extensive experiments on three benchmarks show RDE's robustness and superiority over existing methods.
- Introduction to Text-to-image person re-identification.
- Problem of noisy correspondence in training data.
- Proposal of Robust Dual Embedding method (RDE).
- Components: Confident Consensus Division (CCD) and Triplet Alignment Loss (TAL).
- Experiments on three public benchmarks showcasing RDE's performance.
Stats
Our RDE achieves 75.94%, 90.14%, and 94.12% in terms of Rank-1,5,10 on the 'Best' rows under 0% noise.
The proposed TAL outperforms widely-used Triplet Ranking Loss (TRL) and SDM loss [24].
On CUHK-PEDES with 50% noise, RDE achieves 71.33%, 87.41%, and 91.81% in terms of Rank-1,5,10 on the 'Best' rows.
Quotes
"The model does not know which pairs are noisy in practice."
"Our RDE can achieve robustness against NC thanks to the proposed reliable supervision and stable triplet loss."