Source-linked AI summary

The Devil is in the Details: Self-Supervised Attention for Vehicle Re-Identification

Pirazh Khorramshahi, Neehar Peri, Jun-cheng Chen, Rama Chellappa

arXiv:2004.06271v3cs.CV

TL;DR

Vehicle re-identification methods often require costly annotations or auxiliary information to focus on discriminative regions, limiting scalability across datasets. SAVER learns attention through self-supervised residual generation and deep feature extraction without extra annotations, and outperforms existing state-of-the-art methods across several benchmarks, with limitations on extremely similar images.

  • Problem

    Vehicle re-identification is challenging for same-looking vehicles and viewpoint changes, while annotation-dependent attention methods are costly to scale.

  • Method

    SAVER reconstructs coarse vehicle structure, extracts residual details, and combines residual and input images for feature extraction without extra annotations or attributes.

  • Results

    SAVER outperforms existing state-of-the-art methods on VeRi, VehicleID, Vehicle-1M, and VERI-Wild benchmarks.

  • Takeaways & Limitations

    The results show that attending to discriminative vehicle regions can be learned without additional annotations.

  • Takeaways & Limitations

    On PKU VD1&2, performance is comparable to the baseline and saturated because some images are nearly impossible to distinguish using visual information alone.

Abstract

from arXiv · show

In recent years, the research community has approached the problem of vehicle re-identification (re-id) with attention-based models, specifically focusing on regions of a vehicle containing discriminative information. These re-id methods rely on expensive key-point labels, part annotations, and additional attributes including vehicle make, model, and color. Given the large number of vehicle re-id datasets with various levels of annotations, strongly-supervised methods are unable to scale across different domains. In this paper, we present Self-supervised Attention for Vehicle Re-identification (SAVER), a novel approach to effectively learn vehicle-specific discriminative features. Through extensive experimentation, we show that SAVER improves upon the state-of-the-art on challenging VeRi, VehicleID, Vehicle-1M and VERI-Wild datasets.

1 Introduction

Vehicle re-identification is difficult because visually similar vehicles vary across viewpoints, while discriminative attention methods often depend on costly annotations that do not scale. The paper motivates a scalable alternative for learning vehicle-specific details.

  • Vehicles with different identities can share the same make, model, and color, while appearance changes substantially across viewpoints.
  • Recent methods focus on discriminative regions to improve robustness to orientation and occlusion.
  • High-performing approaches use key-point and part-box annotations alongside object identities to learn global and local representations.
  • These extra annotations improve performance but are costly to collect and difficult to scale as vehicle designs and datasets grow.

2 Related Works

Vehicle re-identification research uses deep embeddings, auxiliary information, and synthetic data to address visually similar vehicles and viewpoint variation. SAVER instead generates attention without extra annotations, attributes, or multimodal information.

  • Global deep representations often fail to distinguish vehicles sharing the same make, model, and color.
  • Prior work augments vehicle re-identification with synthetic images, spatio-temporal cues, license plates, trajectories, camera locations, and timestamps.
  • SAVER uses self-supervised attention generation without extra annotations, attributes, spatio-temporal information, or multimodal information.

3 Self-Supervised Attention for Vehicle Re-identification

SAVER reconstructs a coarse vehicle template, extracts residual details, and combines the residual with the original image for re-identification. Its VAE reconstruction and ResNet-50 extractor are trained through self-supervised and end-to-end objectives.

  • Self-Supervised Residual Generation: The reconstruction network preserves vehicle shape and structure while obfuscating discriminative details, enabling residual-based highlighting of salient regions.
  • Self-Supervised Residual Generation: The VAE downsamples images into a latent representation, applies reparameterization, and upsamples features to produce a coarse reconstruction.
  • Self-Supervised Residual Generation: The reconstruction model is pretrained with mean squared error and Kullback–Leibler divergence objectives.
  • Deep Feature Extraction: The feature extractor uses a single-branch ResNet-50 trained with triplet and cross-entropy classification losses.
  • End-To-End Training: Residuals are combined with the original input as Ic = α × Io + (1 − α) × Ir using a learnable parameter α before ResNet-50 feature extraction.
  • End-To-End Training: End-to-end training jointly optimizes the VAE and feature extractor so residual generation adapts to the re-identification task.

4 Experiments

The experiments section introduces the evaluation datasets, general vehicle re-identification metrics, implementation details, and reported results for SAVER.

  • The evaluation covers dataset descriptions, re-identification evaluation procedures, implementation details, and experimental results.

4.1 Vehicle Re-Identification Datasets

SAVER is evaluated across multiple vehicle re-identification benchmarks with varied identities, images, cameras, and evaluation splits. Performance is measured using ranked-gallery metrics, chiefly CMC and mAP.

  • Datasets: SAVER is evaluated on six vehicle re-identification benchmarks, including VeRi, VehicleID, VERI-Wild, Vehicle-1M, and PKU VD1&VD2.Four additional unconstrained vehicle-image datasets support pre-training of the self-supervised reconstruction network.
  • Datasets: The datasets vary in the numbers of unique identities, images, and cameras used for evaluation.VehicleID, VERI-Wild, Vehicle-1M, and VD1&VD2 use small, medium, and large evaluation splits.
  • Evaluation metrics: CMC@K measures whether a matching vehicle appears within the top K ranked gallery results, while mAP averages precision across queries.CMC@1 and CMC@5 are commonly reported re-identification metrics.

4.2 Implementation Details

The implementation standardizes image preprocessing, pre-trains the self-supervised reconstruction module on Vehicle Universe, and trains the end-to-end feature extractor with ResNet-50.

  • Preprocessing: Images are resized to 256 × 256, normalized across RGB channels, and processed with Detectron to reduce background noise.The normalization uses a mean and standard deviation of 0.5 across RGB channels.
  • Self-supervised residual generation: The residual-generation module is pre-trained on Vehicle Universe, assembled from vehicle datasets spanning varied distributions.Vehicle Universe contains data from sources including CompCars, StanfordCars, BoxCars116K, CityFlow, PKU VD1&VD2, Vehicle-1M, VehicleID, VeRi, and VeRi-Wild.
  • Deep feature extraction: The deep feature extractor uses ResNet-50 with an Adam optimizer for 150 end-to-end training epochs.The initial learning rate is 3.5e−5, with scheduled warm-up and decay; the initial α is 0.5 and γ is 0.3.

4.3 Experimental Evaluation

Across challenging vehicle re-identification benchmarks, SAVER generally improves the baseline and state-of-the-art comparisons, while gains are limited on the highly saturated PKU VD datasets.

  • VeRi: 1.4% mAP improvement over the strong baseline is reported on VeRi.The learned convex combination saturates at α = 0.96, using 96% of the original image and 4% of the residual.
  • VehicleID: +1.0% CMC@1 improvement over the baseline is observed across all VehicleID test splits.Grad-CAM visualizations show attention on discriminative regions such as headlights, the hood, and windshield stickers.
  • VERI-Wild: +2.0% mAP and +1.0% CMC@1 improvements over the baseline are reported across all VERI-Wild evaluation splits.The final α = 0.94 indicates greater residual contribution in this dataset.
  • Vehicle-1M: Vehicle-1M performance improves over the baseline in all metrics across all evaluation sets with α = 0.98.The evaluation splits are randomly constructed and kept fixed throughout the experiments.
  • PKU VD1&VD2: SAVER is comparable to the baseline on PKU VD1&VD2, where the reported α = 0.99 indicates little remaining room for improvement.The paper attributes the saturation to the extreme visual similarity of some images, including nearly indistinguishable windshield sticker placement.
  • State-of-the-art comparison: SAVER surpasses recent state-of-the-art vehicle re-identification methods without extra annotations or attributes.With re-ranking, the reported mAP scores are 82.0 on VeRi and 84.4 on VERI-Wild.

5 Ablation Studies

The ablations examine reconstruction architectures, KL-loss scaling, and strategies for incorporating residual information into vehicle re-identification. Results show that reconstruction quality and residual usage must preserve discriminative local details rather than distort or discard them.

  • Reconstruction architectures: AE and GAN reconstructions attempt to recover fine-grained details but introduce distortions, including altered textures, logos, and vehicle shape.Bilateral Filtering is also evaluated as a texture-smoothing baseline.
  • KL-loss scaling: Higher λ values produce blurrier VAE reconstructions, while λ = 1e −3 yields higher-quality vehicle templates but removes discriminative information across all datasets.As λ approaches 0, the VAE approaches the reconstruction quality of a traditional Auto-Encoder.
  • Residual incorporation: Residual-only training performs nearly as well as the standard baseline, reaffirming that local information is essential for vehicle re-identification.Using only VAE reconstructions performs worst because discriminating details are obfuscated.
  • Residual incorporation: Concatenating residual and original images performs below the baseline because ImageNet initialization is not well suited to six-channel inputs.This corresponds to experiment D.

6 Conclusion

The paper presents scalable architectures and training procedures for robust vehicle re-identification features. SAVER attends to discriminative regions without additional annotations and outperforms existing state-of-the-art methods on several benchmark datasets.

  • The approach uses simple, highly scalable network architectures and training procedures to generate robust deep features for vehicle re-identification.
  • SAVER highlights discriminative regions without requiring additional annotations.
  • SAVER outperforms existing state-of-the-art methods on VeRi, VehicleID, Vehicle-1M, and VeRi-Wild.
Loading 2004.06271v3…