FPAN: Fine-grained and progressive attention localization network for data retrieval

Sijia Chen, Bin Song, Jie Guo, Yanling Zhang, Xiaojiang Du, Mohsen Guizani

Abstract

Automated object localization in massive visual datasets is a critical requirement for modern Intelligent and Connected Transportation Systems (ICTS). This paper introduces FPAN (Fine-grained and Progressive Attention Localization Network), an end-to-end framework that locates queried objects within complex images. By integrating a progressive coarse-to-fine attention mechanism with a learnable cascade up-sampling module, FPAN effectively suppresses background noise while preserving high-resolution spatial details. Experimental results on synthetic retrieval datasets and standard tracking benchmarks demonstrate that FPAN achieves state-of-the-art localization precision and real-time performance (40 FPS), significantly outperforming traditional template-matching and deep learning baselines.

Introduction

The rapid growth of visual data in transportation hubs necessitates advanced retrieval systems capable of identifying specific vehicles or individuals across thousands of camera feeds. Traditional manual search is unscalable, while classic template-matching algorithms often fail in the presence of complex backgrounds, occlusions, or object deformations.

The data retrieval workflow using intelligent localization technology to find specific targets.

Recent advancements in deep learning have utilized visual attention mechanisms to mimic human search patterns. However, standard deep architectures suffer from an inherent trade-off: as the network deepens to capture semantic meaning, spatial resolution is lost, leading to imprecise localization. FPAN addresses this by fusing deep semantic features with shallow spatial information in a unified, fully convolutional architecture.

Methodology

FPAN treats object localization as a multi-task learning problem, processing a target image and a query patch simultaneously. Its architecture is defined by three core innovations:

Progressive Attention Mechanism

The network employs a series of $N$ blocks that progressively refine the search area. As data flows through the layers, the network learns to suppress irrelevant background clutter, focusing its "attention" on the target in a coarse-to-fine manner.

Fine-Grained Attention via CSCONV

To generate high-quality attention maps, the authors introduced Channel-Separated Convolution (CSCONV). Inspired by depth-wise separable convolutions, CSCONV extracts multi-scale local features with high efficiency. These features are then fused into an ensemble of attention maps, enhancing the model's robustness against noise.

Top-Down Cascade Up-sampling

To solve the resolution loss problem, FPAN implements a learnable cascade up-sampling module. This module fuses high-level, low-resolution attention maps from deep layers with high-resolution maps from earlier layers. Through successive deconvolution operations, the network recovers precise spatial details, enabling the generation of accurate bounding boxes.

The overall architecture of FPAN, illustrating the flow from input query to the final localized output.

Experimental Results

FPAN was evaluated across synthetic retrieval tasks and standard tracking benchmarks to assess its accuracy and speed.

  • Retrieval Performance: On the MNIST-RQ dataset—a challenging benchmark featuring digits over cluttered real-world backgrounds—FPAN achieved an Average IOU (AIOU) of 0.71 and an Average Localization Precision (ALP) of 70%. This represents a significant 47% precision gain over traditional baselines.
  • Tracking Benchmarks: When adapted for object tracking, FPAN achieved an Expected Average Overlap (EAO) of 0.296 on the VOT-2016 dataset, surpassing established trackers such as MDNet and C-COT.
  • Computational Efficiency: As a unified network, FPAN requires only a single forward pass per query. It maintains a high frame rate of approximately 40 FPS, making it 3 to 20 times faster than competing methods.

Conclusion

FPAN provides a powerful framework for real-time object localization in complex visual environments. By combining progressive attention with a resolution-recovering up-sampling strategy, it bridges the gap between semantic understanding and spatial precision. Future research will focus on improving robustness to heavy occlusion through the integration of temporal memory structures.

Summary updated

BibTeX
@article{chen2018fpan,
  author = {Sijia Chen and Bin Song and Jie Guo and Yanling Zhang and Xiaojiang Du and Mohsen Guizani},
  title = {{FPAN: Fine-grained and progressive attention localization network for data retrieval}},
  journal = {Computer Networks},
  year = {2018},
  url = {https://openreview.net/forum?id=2XF7ahD4ue},
}
Figure