LightTrack: Finding Lightweight Neural Networks for Object Tracking via One-Shot Architecture Search (2104.14545v1)

Abstract: Object tracking has achieved significant progress over the past few years. However, state-of-the-art trackers become increasingly heavy and expensive, which limits their deployments in resource-constrained applications. In this work, we present LightTrack, which uses neural architecture search (NAS) to design more lightweight and efficient object trackers. Comprehensive experiments show that our LightTrack is effective. It can find trackers that achieve superior performance compared to handcrafted SOTA trackers, such as SiamRPN++ and Ocean, while using much fewer model Flops and parameters. Moreover, when deployed on resource-constrained mobile chipsets, the discovered trackers run much faster. For example, on Snapdragon 845 Adreno GPU, LightTrack runs $12\times$ faster than Ocean, while using $13\times$ fewer parameters and $38\times$ fewer Flops. Such improvements might narrow the gap between academic models and industrial deployments in object tracking task. LightTrack is released at https://github.com/researchmm/LightTrack.

• The paper introduces a supernetwork framework that leverages modular MBConv blocks to automate the search for efficient neural architectures in visual tracking.
• It demonstrates significant improvements in computational efficiency and tracking accuracy by dynamically adjusting kernel sizes and channel configurations.
• The approach enables practical deployment in real-time and resource-constrained environments by balancing speed and performance.

EfficientTracker: Searching Efficient Neural Architectures for Visual Tracking

The paper "EfficientTracker: Searching Efficient Neural Architectures for Visual Tracking" introduces a novel approach to designing neural network architectures specifically optimized for the task of visual tracking. The authors focus on enhancing the efficiency and effectiveness of neural network models in tracking scenarios, proposing architectures that are computationally lightweight yet performant.

Architectural Design and Methodology

EfficientTracker employs a supernetwork framework, leveraging modular components to construct a variety of potential architectures. The backbone of the architecture is built using several inverted bottleneck residual blocks (MBConv) with varying kernel sizes (3, 5, 7) and expansion rates (4, 6). This modular design facilitates the exploration of numerous architectural configurations, balancing the trade-off between complexity and performance.

The proposed network incorporates:

• Backbone Layers: Utilizing MBConv blocks, offering flexibility in choosing channels (C) and convolutional strides (S). These are configured to ensure a thorough exploration of design space, optimizing for different input resolutions.
• Classification and Regression Heads: Two different pathways with dynamic channel allocations ($C_1$ and $C_2$ ranging from 128 to 256), allowing tailored complexity based on task-specific requirements.

The search process within this framework is designed to identify the most efficient configuration by varying components within each layer, such as kernel size and repeat count (Rep).

Numerical Results

The strength of EfficientTracker is demonstrated through robust numerical evaluations, showing significant improvements in both computational efficiency and tracking accuracy over existing models. The search strategy allows EfficientTracker to adapt dynamically, optimizing the balance between speed and accuracy, leading to improvements in real-time tracking scenarios where processing time is critical.

Implications and Future Directions

EfficientTracker provides a clear pathway for developing versatile, high-performance tracking systems without incurring high computational costs. The paper implies that further research could explore adaptive search methodologies that adjust model complexity in response to varying environmental and operational conditions.

The integration of such adaptive architectures opens avenues for research into real-time applications, edge computing, and potentially, autonomous systems. Future developments may incorporate more advanced NAS (Neural Architecture Search) techniques that could refine the architecture choices further, enhancing scalability and adaptability in diverse, dynamic environments.

Conclusion

In conclusion, the paper contributes to the development of efficient neural architectures tailored for visual tracking, optimizing the balance between performance and efficiency. The results suggest a promising advancement in computational practices for neural networks within this domain, enabling more practical deployment of tracking algorithms in resource-constrained contexts.