[Paper Review] Relation Preserving Triplet Mining for Stabilizing the Triplet Loss in Vehicle Re-identification.

Adhiraj Ghosh, Kuruparan Shanmugalingam+1 more|arXiv (Cornell University)|Oct 15, 2021

Video Surveillance and Tracking Methods42 references4 citations

TL;DR

This paper proposes Relation Preserving Triplet Mining (RPTM), a novel triplet mining strategy that respects natural subgroups within vehicle re-identification instances caused by pose variations. By preserving intrinsic relations among samples, RPTM stabilizes the triplet loss and enables a single network with fixed parameters to achieve state-of-the-art performance across three major benchmarks.

ABSTRACT

Object appearances often change dramatically with pose variations. This creates a challenge for embedding schemes that seek to map instances with the same object ID to locations that are as close as possible. This issue becomes significantly heightened in complex computer vision tasks such as re-identification(re-id). In this paper, we suggest these dramatic appearance changes are indications that an object ID is composed of multiple natural groups and it is counter-productive to forcefully map instances from different groups to a common location. This leads us to introduce Relation Preserving Triplet Mining (RPTM), a feature matching guided triplet mining scheme, that ensures triplets will respect the natural sub-groupings within an object ID. We use this triplet mining mechanism to establish a pose-aware, well-conditioned triplet cost function. This allows a single network to be trained with fixed parameters across three challenging benchmarks, while still providing state-of-the-art re-identification results.

Motivation & Objective

• To address the challenge of dramatic appearance variations in vehicle re-identification due to pose changes.
• To identify that forcing all instances of the same ID into a single embedding region is counter-productive when natural subgroups exist.
• To develop a triplet mining mechanism that respects intrinsic sub-groupings within the same ID to improve embedding quality.
• To design a pose-aware, well-conditioned triplet loss function that enables stable training with fixed parameters across benchmarks.

Proposed method

• Introduce Relation Preserving Triplet Mining (RPTM), a feature matching-guided triplet mining scheme that identifies and respects natural subgroups within the same object ID.
• Use the structure of appearance variations due to pose to guide triplet selection, ensuring positive samples are drawn from the same natural subgroup.
• Formulate a pose-aware triplet loss function that maintains relative distances within subgroups while preserving inter-ID discrimination.
• Train a single deep network with fixed hyperparameters across three benchmark datasets using the RPTM-guided loss.
• Ensure the triplet mining process avoids pulling apart instances from different subgroups of the same ID, reducing embedding collapse.

Experimental results

Research questions

• RQ1 How can triplet mining be improved to handle dramatic appearance variations caused by pose changes in vehicle re-identification?
• RQ2 To what extent do natural subgroups within the same vehicle ID affect the stability and performance of triplet loss?
• RQ3 Can a relation-preserving mining strategy lead to better generalization across multiple benchmarks with fixed network parameters?
• RQ4 How does preserving intra-ID sub-group structure improve the conditioning of the triplet loss function?

Key findings

• RPTM enables a single deep network to achieve state-of-the-art performance on three major vehicle re-identification benchmarks using fixed hyperparameters.
• The method stabilizes the triplet loss by respecting natural subgroups formed due to pose variations, reducing embedding collapse.
• By avoiding forced merging of distinct subgroups into a single embedding region, the model maintains better intra-ID compactness.
• The pose-aware, relation-preserving mining leads to a well-conditioned loss function that generalizes effectively across diverse datasets.