Abstract:Over the last decade, Siamese network architectures have emerged as dominating tracking paradigms, which have led to significant progress. These architectures are made up of a backbone network and a head network. The backbone network comprises two identical feature extraction sub-branches, one for the target template and one for the search candidate. The head network takes both the template and candidate features as inputs and produces a local similarity score for the target object in each location of the search candidate. Despite promising results that have been attained in visual tracking, challenges persist in developing efficient and lightweight models due to the inherent complexity of the task. Specifically, manually designed tracking models that rely heavily on the knowledge and experience of relevant experts are lacking. In addition, the existing tracking approaches achieve excellent performance at the cost of large numbers of parameters and vast amounts of computations. A novel Siamese tracking approach called TrackNAS based on neural architecture search is proposed to reduce the complexity of the neural architecture applied in visual tracking. First, according to the principle of the Siamese network, backbone and head network search spaces are constructed, constituting the search space for the network architecture. Next, under the given resource constraints, the network architecture that meets the tracking performance requirements is obtained by optimizing a hybrid search strategy that combines distributed and joint approaches. Then, an evolutionary method is used to lighten the network architecture obtained from the search phase to facilitate deployment to devices with resource constraints (FLOPs). Finally, to verify the performance of TrackNAS, comparison and ablation experiments are conducted using several large-scale visual tracking benchmark datasets, such as OTB100, VOT2018, UAV123, LaSOT, and GOT-10k. The results indicate that the proposed TrackNAS achieves competitive performance in terms of accuracy and robustness, and the number of network parameters and computation volume are far smaller than those of other advanced Siamese trackers, meeting the requirements for lightweight deployment to resource-constrained devices.

Visual Tracking Based on Stacked Denoising Autoencoder Network with Genetic Algorithm Optimization.

A low-complexity visual tracking approach with single hidden layer neural networks

An ensemble of deep neural networks for object tracking

Robust Visual Tracking Using Dynamic Classifier Selection with Sparse Representation of Label Noise

Adaptive Image Denoising Based on Improved Stacked Sparse Denoising Auto-Encoder

Discriminative Deep Non-Linear Dictionary Learning for Visual Object Tracking

Image denoising based on improved stacked sparse denoising autoencoder

Visual Object Tracking in RGB-D Data via Genetic Feature Learning

A Comparative Study Of Object Tracking Using Cnn And Sdae

Dual Deep Network for Visual Tracking

Wasserstein Distance-Based Auto-Encoder Tracking.

Enhancement Of Elda Tracker Based On Cnn Features And Adaptive Model Update

Learning a Deep Compact Image Representation for Visual Tracking

Decontaminate Feature for Tracking: Adaptive Tracking Via Evolutionary Feature Subset.

A Real-Time Visual Tracking Approach Using Sparse Autoencoder and Extreme Learning Machine

Learning Cascaded Siamese Networks for High Performance Visual Tracking

Denoising Auto-Encoders Toward Robust Unsupervised Feature Representation.

A deep features based generative model for visual tracking

High-precision and Real-Time Visual Tracking Algorithm Based on the Siamese Network for Autonomous Driving

Efficient and Lightweight Visual Tracking with Differentiable Neural Architecture Search

Learning Dynamic Siamese Network For Visual Object Tracking