Global Dilated Attention and Target Focusing Network for Robust Tracking

Yun Liang, Qiaoqiao Li, Fumian LongAAAI 2023 · DOI

TL;DR

This paper introduces GdaTFT, a novel visual tracking framework that enhances semantic understanding and target focus. By employing a Global Dilated Attention (GDA) module for improved feature representation and a Target Focusing Network (TFN) for precise target localization, GdaTFT achieves state-of-the-art performance on challenging benchmarks, outperforming existing methods in robustness and accuracy.

What Problem Does This Paper Address?

Visual tracking, the task of predicting a target's state in a video sequence, is crucial for applications like surveillance and autonomous driving. Despite advancements, trackers struggle with challenges like rapid deformations, occlusions, and cluttered backgrounds. While Siamese network-based trackers have gained popularity for their balance of accuracy and speed, they face limitations:

  1. Limited Semantic Understanding: Their fixed convolutional operations and small receptive fields hinder the extraction of rich semantic information, making it difficult to distinguish targets from similar-looking backgrounds or distractors.
  2. Feature Fusion Issues: Existing methods often fuse features linearly, which can introduce noise and dilute the target's representation, especially when dealing with semantically similar non-target objects.

This paper aims to address these issues by developing a tracker that can effectively model global semantics and precisely focus on the target, even in the presence of challenging visual conditions.

Methods

The proposed GdaTFT framework consists of three main components: a Siamese backbone for feature extraction, a Global Target Focusing module, and a prediction head.

  • Feature Extraction: A standard Siamese backbone network with shared weights is used to extract features from both the template (initial target patch) and the search region (current frame).
  • Global Target Focusing (GDA + TFN): This is the core innovation.
  • Global Dilated Attention (GDA): This module enhances feature semantics in both spatial and channel dimensions. It comprises:
  • Local Focusing Module (LFM): Uses depth-wise convolution to capture local structural information and dependencies, preserving fine details.
  • Dilated Attention (DA): Employs dilated sampling to create sparse, long-range dependencies within feature maps, allowing for better semantic understanding and reducing background noise. It uses multi-head self-attention within independent groups of sampled features.
  • Channel Adaptation Module (CAM): Utilizes 1x1 convolutions to build inter-channel associations, adaptively strengthening semantically rich channels.
  • Target Focusing Network (TFN): This module further refines the target's representation by building point-to-point associations between the template and search region features. Unlike traditional feature fusion, TFN uses the template as a query to guide the search region features, effectively augmenting target embedding while suppressing non-target semantic objects. This is achieved by calculating a similarity matrix and using it to weight the search region features.
  • Prediction Head: A cross-correlation-based prediction head, common in Siamese trackers, is used to generate a response map for target localization, followed by classification and regression branches.

Datasets

The paper uses the following datasets for training and evaluation:

  • Training: COCO, GOT-10k, ImageNet DET/VID, TrackingNet, and LaSOT.
  • Evaluation: GOT-10k, OTB100, TrackingNet, and LaSOT.

Key Insights and Ideas

  • Dilated Attention for Sparse Long-Range Dependencies: The key idea behind GDA is that by using dilated sampling, the model can capture long-range semantic relationships without the computational burden or noise sensitivity of dense self-attention. This allows for a more robust semantic understanding.
  • Template as Query for Focused Enhancement: TFN's novel approach of using the template as a query is crucial. It allows the target's rich semantic information to directly guide the enhancement of the search region features, leading to a more precise focus on the target and better suppression of distractors.
  • Complementary Nature of GDA and TFN: GDA first enhances global semantics and reduces background interference, creating a cleaner feature space. TFN then leverages this enhanced feature to specifically focus on the target, distinguishing it from other semantically similar objects. This progressive approach is more effective than single-stage methods.
  • Importance of Local and Global Information: The combination of LFM (local details) and DA (global semantics) within GDA ensures that both fine-grained object features and broader contextual information are captured.

Key Results

  • State-of-the-Art Performance: GdaTFT achieves leading performance on challenging benchmarks, including GOT-10k (SR0.75 of 77.8, surpassing TransT), OTB100 (top-tier performance across 6 challenges), and LaSOT (outperforming most trackers and on par with TransT).
  • Improved Robustness: Qualitative results (Figure 7) show GdaTFT's ability to produce more accurate features, effectively distinguishing targets from background and distractors, unlike feature fusion methods that can be confused by similar objects.
  • Effectiveness of GDA and TFN: Ablation studies (Table 2) demonstrate that both GDA and TFN individually improve performance over baseline self-attention and feature fusion methods, respectively. When used together, they yield the best results, confirming their synergistic effect.

Limitations and Future Work

  • Computational Cost: While Dilated Attention aims to be more efficient than dense self-attention, the overall computational cost of the proposed method is not explicitly detailed in terms of speed.
  • Future Work: The paper focuses on improving semantic understanding and target focusing. Future work could explore further optimizations for real-time performance or investigate more complex tracking scenarios like multi-target tracking. The authors also mention making their code available, which will facilitate further research and development in this area.

Summary updated

BibTeX
@inproceedings{liang2023global,
  author = {Yun Liang and Qiaoqiao Li and Fumian Long},
  title = {{Global Dilated Attention and Target Focusing Network for Robust Tracking}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i2.25241},
}
Figure