[Paper Review] Beyond Triplet Loss: Person Re-identification with Fine-grained Difference-aware Pairwise Loss

Yan Cheng, Guansong Pang+3 more|arXiv (Cornell University)|Sep 22, 2020

Video Surveillance and Tracking Methods47 references19 citations

TL;DR

This paper proposes a novel fine-grained difference-aware (FIDI) pairwise loss for person re-identification that adaptively applies exponential penalization to small appearance differences and bounded penalization to large differences, significantly improving feature learning for fine-grained distinctions. The FIDI loss outperforms triplet loss and other state-of-the-art losses across four benchmark datasets, achieving higher mAP and R-1 scores while enhancing data efficiency.

ABSTRACT

Person Re-IDentification (ReID) aims at re-identifying persons from different viewpoints across multiple cameras. Capturing the fine-grained appearance differences is often the key to accurate person ReID, because many identities can be differentiated only when looking into these fine-grained differences. However, most state-of-the-art person ReID approaches, typically driven by a triplet loss, fail to effectively learn the fine-grained features as they are focused more on differentiating large appearance differences. To address this issue, we introduce a novel pairwise loss function that enables ReID models to learn the fine-grained features by adaptively enforcing an exponential penalization on the images of small differences and a bounded penalization on the images of large differences. The proposed loss is generic and can be used as a plugin to replace the triplet loss to significantly enhance different types of state-of-the-art approaches. Experimental results on four benchmark datasets show that the proposed loss substantially outperforms a number of popular loss functions by large margins; and it also enables significantly improved data efficiency.

Motivation & Objective

• To address the limitation of triplet loss in capturing fine-grained appearance differences in person re-identification.
• To develop a loss function that effectively penalizes small inter- and intra-person differences, which are critical for accurate ReID.
• To improve model performance and data efficiency by learning more discriminative representations through adaptive penalization.
• To create a generic, plug-in loss that enhances various state-of-the-art ReID models without architectural changes.

Proposed method

• Proposes a fine-grained difference-aware (FIDI) pairwise loss that applies exponential penalization to small pairwise distances (indicating fine-grained differences) and bounded penalization to large distances.
• The loss function is designed to be sensitive to small differences while avoiding over-penalization on large differences, balancing feature learning across all scales.
• The FIDI loss is formulated as a differentiable function of pairwise embeddings, enabling end-to-end training with standard deep learning frameworks.
• The loss is compatible with any backbone network and can replace triplet loss as a plug-in component in existing ReID models.
• Hyperparameters α (loss bound) and β (sensitivity to distance) are tuned via cross-validation to optimize performance on different datasets.
• The method is evaluated on person ReID benchmarks and generalized to vehicle ReID, demonstrating transferability.

Experimental results

Research questions

• RQ1 Can a pairwise loss that adaptively penalizes small appearance differences improve person ReID performance beyond triplet loss?
• RQ2 How does the FIDI loss enhance feature learning for fine-grained distinctions in complex ReID scenarios?
• RQ3 To what extent does the FIDI loss improve data efficiency compared to triplet loss and other state-of-the-art losses?
• RQ4 Can the FIDI loss generalize effectively to other ReID tasks, such as vehicle ReID?

Key findings

• The FIDI loss achieves 1.3% to 2.6% higher mAP and 0.6% to 0.7% higher R-1 on vehicle ReID datasets compared to the Baseline2 method.
• On the CUHK03-L and CUHK03-D datasets, the FIDI loss improves mAP by over 2% compared to triplet loss-based models.
• The FIDI loss significantly enhances data efficiency, enabling better performance with fewer training samples.
• Visualizations show that models trained with FIDI loss correctly rank images of the same identity even under fine-grained variations, unlike baseline models that confuse similar-looking vehicles.
• The FIDI loss is robust to hyperparameter settings, with optimal performance achieved at α=1.05 and β=0.5, though performance remains stable across a range of values.
• The FIDI loss generalizes well to vehicle ReID, outperforming multiple state-of-the-art methods on VeRi-776 and VehicleID datasets.