PaperlayerView exact revision on arXiv

Source-linked AI summary

Vehicle Re-Identification: an Efficient Baseline Using Triplet Embedding

Ratnesh Kumar, Edwin Weill, Farzin Aghdasi, Parthsarathy Sriram

arXiv:1901.01015v4cs.CV

TL;DR

Vehicle re-identification must match the same vehicle across cameras despite appearance variation and difficult visual distinctions. The paper applies deep triplet-embedding practices and evaluates batch sample; its baselines outperform almost all compared approaches across three datasets, with batch sample and batch weighted generally more effective and robust than batch hard and batch all.

  • Problem

    Vehicle re-identification seeks to identify the same vehicle across cameras, where appearance changes and subtle differences between similar vehicles complicate matching.

    Source: §1 · pp. 1–2

  • Method

    The paper proposes a vehicle re-identification baseline using deep triplet-embedding practices and formally evaluates the batch sample sampling variant.

    Source: §6 · p. 9

  • Results

    Across three datasets, the baselines outperform almost all state-of-the-art approaches across a wide range of evaluation criteria.

    Source: §6 · p. 9

  • Takeaways & Limitations

    Batch sample and batch weighted are generally more effective and robust than batch hard and batch all.

    Source: §6 · p. 9

    Without spatio-temporal information, vehicles of the same color and model with few or no distinguishing marks are very difficult to distinguish.

    Source: §5.6 · p. 9

Abstract

from arXiv · show

In this paper we tackle the problem of vehicle re-identification in a camera network utilizing triplet embeddings. Re-identification is the problem of matching appearances of objects across different cameras. With the proliferation of surveillance cameras enabling smart and safer cities, there is an ever-increasing need to re-identify vehicles across cameras. Typical challenges arising in smart city scenarios include variations of viewpoints, illumination and self occlusions. Most successful approaches for re-identification involve (deep) learning an embedding space such that the vehicles of same identities are projected closer to one another, compared to the vehicles representing different identities. Popular loss functions for learning an embedding (space) include contrastive or triplet loss. In this paper we provide an extensive evaluation of these losses applied to vehicle re-identification and demonstrate that using the best practices for learning embeddings outperform most of the previous approaches proposed in the vehicle re-identification literature. Compared to most existing state-of-the-art approaches, our approach is simpler and more straightforward for training utilizing only identity-level annotations, along with one of the smallest published embedding dimensions for efficient inference. Furthermore in this work we introduce a formal evaluation of a triplet sampling variant (batch sample) into the re-identification literature.

1. Introduction

Vehicle re-identification seeks to identify the same vehicle across cameras, despite viewpoint, lighting, and pose changes that can make visual matching difficult. The paper evaluates embedding-learning practices for this task and formally evaluates the batch sample triplet-sampling variant.

1. Introduction

  • Vehicle re-identification is practically valuable because license plates may be invisible in street-camera views, while vehicles vary in appearance, viewpoint, lighting, and pose.Distinguishing same-color, same-model vehicles may require subtle cues such as dust, writing, or dents.Source: §1 · pp. 1–2
  • Vehicles of the same identity can look diverse across views, yet share color and model; vehicle labels are also less fine-grained than person or face identities.Finite color and model categories mean less dataset diversity than person or face re-identification datasets.Source: §1 · p. 2
  • The paper evaluates informative sampling methods for vehicle re-identification and reports state-of-the-art performance across diverse datasets and metrics.The evaluation applies recent sampling advances from person re-identification to the vehicle task.Source: §1 · p. 2
  • The paper formally evaluates batch sample, a triplet-sampling variant, for re-identification.The paper notes that batch sample was discussed in an implementation webpage but lacked a formalized study and evaluation, as far as the authors knew.Source: §1 · p. 2

2. Related Works

Related work covers CNN-based embeddings trained with contrastive or triplet losses, softmax-based feature learning, and vehicle-specific retrieval methods. These approaches address vehicle appearance variation through ranking, attributes, view synthesis, and spatio-temporal information.

2. Related Works

  • CNN feature-learning methods surpassed handcrafted-feature baselines, while triplet loss uses three samples and contrastive loss uses two; both have reported state-of-the-art performance.Triplet loss is more computationally expensive, with its additional context attributed as a reason for its performance.Source: §2 · p. 2
  • A softmax-trained embedding can be obtained by removing the classification layer, but classification features may not distinguish unseen identities and training can become costly at very large identity counts.The cited example considers datasets with 10^6 identities.Source: §2 · p. 2
  • Vehicle re-identification must separate vehicles within the same model category, unlike model classification; prior methods combine visual features, candidate refinement, and sometimes license-plate or spatio-temporal cues.The cited approaches include handcrafted color and texture features with CNN attributes, and candidate pruning using license-plate matching.Source: §2 · pp. 2–3
  • Prior vehicle methods use model classification with ranking [9], class-center triplet anchors, or multi-grain attributes and ranking.Source: §2 · p. 3
  • Other approaches address intra-class variation with sub-clustering, synthesize or attend to multiple views, or use keypoints and camera-transition information [48] [40].The cited methods combine view synthesis with contrastive, reconstruction, and adversarial losses, or add spatio-temporal regularization during retrieval [48].Source: §2 · p. 3

3. Loss functions for embedding

Embedding learning seeks to place same-identity images close together while remaining robust to changes in viewpoint, illumination, and shape, with efficient computation important for deployment. The section contrasts contrastive and triplet objectives and examines sampling methods for making their training practical.

3. Loss functions for embedding

  • The embedding maps images into a feature space where same-identity images are metrically close; desired properties include invariance to viewpoint, illumination, and shape changes.Efficient embedding computation and ranking are also identified as practical deployment requirements.Source: §3 · p. 3

3.1. Contrastive Loss

  • Contrastive loss pulls same-class distances toward zero and keeps different-class distances above a shared threshold α, an assumption that restricts the embedding manifold.The section also gives its training complexity as O(N^2) for N samples.Source: §3.1 · p. 3

3.2. Triplet Loss

  • Triplet loss compares positive and negative distances from the same anchor, forcing same-class points closer than points from another class.Its O(N^3) training complexity is described as computationally prohibitive, motivating sampling approaches.Source: §3.2 · p. 4

Dataset Sampling

  • Hard mining presents poorly performing examples to the optimizer, but outliers among hard examples can impair discrimination; semihard sampling seeks moderate triplets, though its offline CPU generation impedes convergence.[38] proposed semihard sampling, and [11] proposed mining samples directly on GPU.Source: Dataset Sampling · p. 4
  • Batch sample (BS) is formally evaluated as a re-identification sampling variant; it mines positive and negative examples from an anchor-to-sample distance distribution.The paper says BS had been discussed in [11] but, to the authors’ knowledge, had not received a formalized study and evaluation in the re-identification literature.Source: Dataset Sampling · p. 4
  • Batch hard (BH) selects the hardest positive and negative per anchor, while batch all (BA) weights samples uniformly and batch weighted (BW) emphasizes harder samples by distance.Uniform weighting may underweight informative tough samples because trivial easy samples typically outnumber them.Source: Dataset Sampling · p. 4

4. Experiments

The experiments specify three vehicle datasets, the training setup, and the retrieval metrics used to evaluate the approaches.

4. Experiments

  • Experiments use VeRi, VehicleID, and PKU-VD, spanning different dataset sizes and camera-viewpoint settings.VeRi includes varied viewpoints; VehicleID contains front and rear views, while PKU-VD includes high-resolution and surveillance-camera images.Source: §4 · pp. 4–5

4.1. Training and Hyperparameters

  • Training uses MobileNet, Adam, 128-dimensional embeddings, and softplus margins for triplet loss but a 1.0 hard margin for contrastive loss.The authors select MobileNet for efficiency; the contrastive-loss softplus variant produced poorer results.Source: §4.1 · p. 5

4.2. Evaluation Metrics

  • Evaluation uses mAP and top-k accuracy; AP(q) uses precision at ranks of correct retrievals, and mAP averages AP across queries.The evaluation retrieves gallery identities for each query vehicle.Source: §4.2 · p. 5

5. Results and Discussions

Across the reported datasets, triplet-embedding configurations achieve strong retrieval results, with batch sampling variants leading on several comparisons. Results also show that performance depends on the loss, sampling setup, and dataset; the VeRi comparison highlights an efficient 128-dimensional embedding without spatio-temporal information.

5.1. VeRi

  • On VeRi, our approach outperforms all compared results on mAP, while GSTE achieves better top-k accuracy, and the mAP advantage indicates robustness at all ranks.GSTE uses a complicated training process requiring tuning an additional intra-class clustering parameter.Source: Table 3 · p. 6, §5.1 · p. 6

5.2. PKU-VD

  • On PKU-VD, using more triplets in a batch improves accuracy, but batch-hard sampling does not start training from scratch; other variants are more robust than default PK batch-size training.The paper notes ImageNet-pretrained initialization as one way to use batch-hard sampling.Source: §5.2 · p. 7, Table 6 · p. 7, Table 7 · p. 7
  • On PKU-VD, BW sampling with batch size 18x16 outperforms the previous state of the art, without vehicle attributes and with a 128-dimensional rather than 1024-dimensional embedding.The paper describes Multi-grain ranking as using vehicle attributes during training.Source: §5.2 · p. 7, Table 7 · p. 7

5.3. VehicleID

  • On VehicleID, larger batches improve retrieval rankings, but by a smaller margin; the authors suggest limited viewpoint and vehicle-model variability may constrain the added informative statistics.The dataset evaluation averages metrics over ten randomly constructed gallery sets for each reference set.Source: §5.3 · pp. 7–8
  • On VehicleID, BS and BW outperform other sampling variants and all compared approaches on mAP, while GSTE leads in top-1 but drops at top-5.OIFE+ [48] has close top-5 accuracy to the approach but uses keypoint annotations and a separate metric-learning module.Source: §5.3 · p. 8, Table 9 · p. 8, Table 10 · p. 8

5.4. Cityflow dataset

  • The CityFlow results are referred to in the dataset authors’ paper, rather than reported in the supplied section passages.Source: §5.4 · p. 8

5.5. VRIC 3

  • On VRIC, the standard experimental settings outperform the dataset authors’ baseline multi-scale matching approach.The table caption specifies ImageNet-pretrained MobileNet-v1, PK batch size 18x4, and 128-dimensional embeddings.Source: §5.5 · p. 8, Table 12 · p. 8

5.6. Veri-Wild3

  • The paper identifies distinguishing vehicles of the same color and model without spatio-temporal information as difficult, and names combining that information with other attributes as future work.The supplied passages provide no comparative outcome for Veri-Wild beyond pointing to its results table.Source: §5.6 · p. 9, Table 13 · p. 9

6. Conclusion and Future Work

The paper presents a triplet-embedding baseline for vehicle re-identification and reports outperforming almost all state-of-the-art approaches across three datasets. Batch sample and batch weighted were generally more effective and robust than batch hard and batch all.

  • The proposed baseline applies deep triplet-embedding best practices, constructing batches to extract meaningful statistics that guide training and convergence.Source: §6 · p. 9
  • The baselines outperform almost all state-of-the-art approaches across three datasets, while batch sample and batch weighted are generally more effective and robust than batch hard and batch all.Source: §6 · p. 9
  • The authors frame vehicle identity as less fine-grained than person re-identification and report that recent embedding advances demonstrate this premise.Source: §6 · p. 9
  • Figure 2 shows top-10 retrievals for VeRi queries, with correct and incorrect results marked, and includes matches across views and cameras.Source: Figure 2 · p. 10
Loading 1901.01015v4…