[Paper Review] Siamese Anchor Proposal Network for High-Speed Aerial Tracking

Changhong Fu, Ziang Cao+3 more|arXiv (Cornell University)|Dec 19, 2020

Video Surveillance and Tracking Methods33 references4 citations

TL;DR

This paper proposes SiamAPN, a two-stage Siamese network for high-speed aerial tracking that uses a novel anchor proposal network (APN) to generate adaptive, no-prior anchors, reducing hyperparameters and improving robustness to small, occluded, and fast-moving objects. The method achieves ~200 FPS with state-of-the-art performance on three UAV benchmarks, outperforming SiamRPN++ and SiamFC++ in precision and AUC across challenging scenarios like fast motion and low resolution.

ABSTRACT

In the domain of visual tracking, most deep learning-based trackers highlight the accuracy but casting aside efficiency. Therefore, their real-world deployment on mobile platforms like the unmanned aerial vehicle (UAV) is impeded. In this work, a novel two-stage Siamese network-based method is proposed for aerial tracking, extit{i.e.}, stage-1 for high-quality anchor proposal generation, stage-2 for refining the anchor proposal. Different from anchor-based methods with numerous pre-defined fixed-sized anchors, our no-prior method can 1) increase the robustness and generalization to different objects with various sizes, especially to small, occluded, and fast-moving objects, under complex scenarios in light of the adaptive anchor generation, 2) make calculation feasible due to the substantial decrease of anchor numbers. In addition, compared to anchor-free methods, our framework has better performance owing to refinement at stage-2. Comprehensive experiments on three benchmarks have proven the superior performance of our approach, with a speed of around 200 frames/s.

Motivation & Objective

• Address the trade-off between efficiency and robustness in deep learning-based aerial tracking for resource-constrained UAV platforms.
• Overcome limitations of anchor-based methods that rely on pre-defined anchors, which are sensitive to hyperparameters and struggle with variable object sizes.
• Improve generalization and reduce computational redundancy compared to anchor-free methods by introducing adaptive anchor generation without prior assumptions.
• Enhance tracking performance under UAV-specific challenges such as fast motion, low resolution, and occlusion through a lightweight, efficient architecture.
• Achieve real-time inference speed (~200 FPS) while maintaining high accuracy on standard aerial tracking benchmarks.

Proposed method

• Propose a Siamese two-stage framework: stage-1 uses an Anchor Proposal Network (APN) to generate one adaptive anchor per feature map point, eliminating fixed anchor priors.
• APN learns to predict adaptive anchor locations and sizes end-to-end, improving generalization to objects of varying sizes and reducing the number of anchors by ~90%.
• Introduce a feature fusion network to integrate spatial location information from adaptive anchors into feature maps, enhancing semantic representation and regression accuracy.
• Leverage a multi-classification structure to improve discriminability and reduce negative sample imbalance during training.
• Design the network for high-speed inference by minimizing redundant anchor proposals and optimizing feature extraction with lightweight components.
• Train the Siamese backbone end-to-end with joint classification and regression heads, using cross-entropy and IoU-based loss functions.

Experimental results

Research questions

• RQ1 Can a Siamese network with adaptive anchor generation achieve high accuracy while maintaining real-time inference speed for aerial tracking?
• RQ2 How does eliminating fixed anchor priors affect robustness to small, occluded, and fast-moving objects in UAV tracking?
• RQ3 To what extent does feature fusion of anchor location information improve tracking performance compared to standard anchor-based or anchor-free methods?
• RQ4 Can the proposed method outperform state-of-the-art trackers on standard UAV benchmarks in terms of precision, AUC, and speed?
• RQ5 Does the reduction in hyperparameters and anchor count lead to improved generalization and training stability in complex aerial scenarios?

Key findings

• SiamAPN achieves ~200 FPS, making it suitable for real-time deployment on mobile and UAV platforms.
• On UAV20L, SiamAPN achieves 76.2% precision and 56.3% AUC, outperforming the second-best tracker (SiamRPN++) by 2.5% in precision and 0.6% in AUC.
• On VisDrone2018-SOT-test, SiamAPN achieves 81.4% precision and 58.4% AUC, ranking first and surpassing UDT+ (78.6%, 57.8%) and SiamFC++ (73.8%, 56.2%).
• On UAV123@10fps, SiamAPN achieves 75.2% precision and 56.6% success rate, ranking first with a 2.3% improvement in precision over SiamRPN++.
• In attribute-based evaluation, SiamAPN improves AUC by 3.9% on average over SiamFC++ in fast motion, 4.0% in full occlusion, and 3.9% in low resolution.
• SiamAPN outperforms SiamRPN++ by 5.1% in AUC on fast motion and 4.55% in full occlusion, demonstrating superior robustness under UAV-specific challenges.