S4-KD: A single step spiking SiamFC+ + for object tracking with knowledge distillation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Neuromorphic Engineering

Background:

• Spiking neural networks (SNNs) offer high efficiency and low energy consumption for information processing via binary spikes.
• Current SNNs often require multiple time steps, leading to increased latency and power usage, hindering real-world applications.
• Optimizing SNNs for tasks like object tracking necessitates reducing temporal complexity without sacrificing accuracy.

Purpose of the Study:

• To develop an efficient, single-time-step Spiking neural network framework for object tracking.
• To enhance the tracking performance of the single-step SNN using knowledge distillation.
• To establish a new state-of-the-art in SNN-based object tracking with improved efficiency and accuracy.

Main Methods:

• Proposed the Single Step Spiking SiamFC++ (S4) framework, compressing temporal steps to one via temporal pruning using AlexNet.
• Introduced a knowledge distillation (KD) approach (S4-KD) using an AlexNet-based ANN as a teacher and the S4 model as a student.
• Designed three distinct distillation loss functions to facilitate knowledge transfer in S4-KD.

Main Results:

• The S4 framework achieved tracking performance comparable to the original multi-step Spiking SiamFC++ with a single time step.
• S4-KD demonstrated superior performance on OTB100, UAV123, and VOT2018 benchmarks, achieving state-of-the-art results.
• S4-KD achieved an estimated energy consumption of only 34.6% compared to the original Spiking SiamFC++.

Conclusions:

• Single-step SNNs are viable for efficient object tracking, maintaining competitive performance.
• Knowledge distillation is an effective strategy for boosting the performance of compressed SNNs.
• The S4-KD tracker represents a significant advancement in SNN-based object tracking, offering state-of-the-art efficiency and accuracy.