[논문 리뷰] GradNet: Gradient-Guided Network for Visual Object Tracking

Peixia Li, Boyu Chen외 4명|arXiv (Cornell University)|2019. 09. 15.

Video Surveillance and Tracking Methods참고 문헌 42인용 수 49

한 줄 요약

GradNet은 Gradients 기반 템플릿 업데이트를 도입하여 시메이즈 기반 시각 추적에서 그라디언트 정보를 사용해 온라인으로 템플릿을 업데이트하고 템플릿 일반화 학습 방법으로 정확도를 높이면서 실시간 속도를 유지합니다.

ABSTRACT

The fully-convolutional siamese network based on template matching has shown great potentials in visual tracking. During testing, the template is fixed with the initial target feature and the performance totally relies on the general matching ability of the siamese network. However, this manner cannot capture the temporal variations of targets or background clutter. In this work, we propose a novel gradient-guided network to exploit the discriminative information in gradients and update the template in the siamese network through feed-forward and backward operations. Our algorithm performs feed-forward and backward operations to exploit the discriminative informaiton in gradients and capture the core attention of the target. To be specific, the algorithm can utilize the information from the gradient to update the template in the current frame. In addition, a template generalization training method is proposed to better use gradient information and avoid overfitting. To our knowledge, this work is the first attempt to exploit the information in the gradient for template update in siamese-based trackers. Extensive experiments on recent benchmarks demonstrate that our method achieves better performance than other state-of-the-art trackers.

연구 동기 및 목표

• Appearance 변화와 배경 잡음에 대응하기 위한 Siamese 기반 트래커의 온라인 적응을 촉진한다.
• 현재 프레임에서 추적 템플릿을 업데이트하기 위해 그라디언트 정보를 활용한다.
• 단 한 번의 역전파로 템플릿을 적응시키는 경량의 엔드투엔드 네트워크(GradNet)를 개발한다.
• 템플릿 일반화 학습 방법을 통해 온라인 과적합을 방지하고 일반화를 개선한다.

제안 방법

• Two-branch architecture: a search-region feature extractor and an update branch that generates a new template from gradients.
• Initial embedding of the target feature to form an initial template via a sub-network U1.
• Compute the gradient of the loss with respect to the target feature and process it through a second sub-network U2 to produce a gradient-based update.
• Update the template by combining the gradient-driven update with the initial target feature, then re-derive an optimal template for final scoring.
• Train the update branch with second-order gradients and a template generalization strategy that uses cross-video search regions to avoid overfitting.
• During online tracking, update the template every few frames and fuse it with the initial template to balance adaptation and stability.

실험 결과

연구 질문

• RQ1 Can gradient information be exploited to update the Siamese tracker template in real time?
• RQ2 Does a template generalization training strategy reduce overfitting and improve generalization across diverse video domains?
• RQ3 How does GradNet compare to online-update trackers and pure offline Siamese trackers in accuracy and speed?

주요 결과

• GradNet은 표준 GPU/CPU 구성에서 실시간 추적을 80 fps로 달성한다.
• The gradient-guided update improves precision and success metrics over the SiameseFC baseline.
• 템플릿 일반화 학습 reduces overfitting and encourages the update branch to rely on discriminative gradients rather than appearance alone.
• A single backward propagation plus two forward passes suffices to update the template effectively, balancing speed and accuracy.
• Ablation studies show each component (gradient use, template generalization, online update) contributes to performance gains.
• On four benchmarks, GradNet provides competitive or superior performance compared with state-of-the-art real-time trackers.