Adaptive Weighted CNN Features Integration for Correlation Filter Tracking

Abstract

Visual object tracking is an active and challenging research topic in computer vision, as objects often undergo significant appearance changes caused by occlusion, deformation, and background clutter. Although convolutional neural network (CNN)-based trackers have achieved competitive results, there are still some limitations. Most existing CNN-based trackers track the object by leveraging high-level semantic features of the highest convolutional layer, which may lead to low-spatial resolution feature maps and degrade the localization precision of tracking. Furthermore, these trackers hardly benefit from end-toend training since the extraction of features and the learning of classifier are separated. To deal with the above-mentioned issues, we design an adaptive weighted CNN features-based Siamese network for tracking. To capture spatial and semantic information of the object, we design a feature extraction network that derives feature maps by concatenating features of all convolutional layers. To make the features representation more discriminative, we propose a feature integration network. In the feature integration network, we propose a holistic-part network to capture strong visual cues and learn the semantic relations between the holistic object and its parts and combine the holistic-part network with spatial and channel attention mechanisms to adaptively assign weights to each region and channel of the feature maps. In addition, the designed Siamese network can be trained offline end-to-end. The experimental results on the benchmark datasets OTB50 and OTB100 demonstrate that the proposed tracker achieves favorable performance against several state-of-theart trackers while running at an average speed of 20.5 frames/s. INDEX TERMS Visual object tracking, correlation filter, Siamese convolutional neural network, feature integration, channel attention and spatial attention.

Efficient Visual Tracking With Stacked Channel-Spatial Attention Learning

“…Therefore, most previous attention models, e.g. [22], [41], [42], and [53], use either global average or max pooling with a multilayer perceptron (MLP) to calculate their gain. In contrast, rather than a single pooling operation, we consider the global average and max pooling together to construct a channel attention module that learns fused features.…”

“…Previously, Qin et al [54] constructed a spatial mask using super-pixels to exploit target representation. Li et al [53] utilized global max pooling to encode the spatial attention in their model. We exploit the relationship among channels inter-spatial features to construct spatial attention.…”

Shape Matching Based on Multi-Scale Invariant Features

“…The global and local features can be obtained by controlling the side length of the triangular. Convolution neural network (CNN) has also been adopted in shape matching and retrieval [30], [31]. Li and Yang [31] designed an adaptive weighted CNN features-based Siamese network.…”

“…Convolution neural network (CNN) has also been adopted in shape matching and retrieval [30], [31]. Li and Yang [31] designed an adaptive weighted CNN features-based Siamese network. In this method, the feature extraction network is designed to derive feature maps while the feature integration network is adopted to adaptively assign weights to each region of the feature maps and integrate feature maps.…”

Channel Compression: Rethinking Information Redundancy Among Channels in CNN Architecture

“…C ONVOLUTIONAL neural networks (CNNs) are attracting considerable attention in an increasing array of area, such as computer vision [1]- [3], computational acoustics [4]- [6] and natural language processing [7]- [9]. The general trend is to design deeper and more complicated network architecture to pursue better performance.…”