Fully Conventional Anchor-Free Siamese Networks for Object Tracking

Abstract

Siamese network have been extensively applied in the tracking field because of its huge speed advantage and great precision performance in solving the tracking problems. In this paper, we propose an efficient framework for real-time object tracking which is end-to-end trained offline-Fully Conventional Anchor-Free Siamese network (FCAF). Specifically, as the backbone network in Siamese trackers is relatively shallow, resulting in insufficient feature information acquired by the trackers and lower accuracy, the deep network ResNet-50 is adopted to provide richer feature representation. Meanwhile, the introduction of multi-layer feature fusion module effectively combines low-level detail information with high-level semantic features, improving the localization performance. In addition, we propose the anchor-free proposal network (AFPN) to replace the region proposal network (RPN). AFPN network consists of correlation section, implemented by depth-wise cross correlation, and supervised section which has two branches, one for classification and the other for regression. In order to suppress the prediction of low quality bounding boxes, center-ness branch is added. We conduct extensive experiments on the OTB2015 and VOT2016 public datasets, demonstrating that our proposed tracker achieves state-of-the-art performance. INDEX TERMS Object tracking, deep Siamese network, feature fusion, anchor-free network.

Cooperative Use of Recurrent Neural Network and Siamese Region Proposal Network for Robust Visual Tracking

“…Convolutional neural networks have been successfully applied to object detection and recognition because of their powerful feature representation capabilities [5], and this has inspired the introduction of deep learning to solve the challenges posed by object tracking [6][7][8]. Although this helps to improve the accuracy of tracking, deep learning-based target tracking algorithms are computationally expensive when extracting deep features or fine-tuning the network online.…”

“…coordinates, length and width of the target bounding box in the dataset, respectively. Then, the normalized distance is expressed as(6).…”

PGNet: Pipeline Guidance for Human Key-Point Detection

“…In the feature extraction process, there are two main methods of feature extraction. One is box-of-free feature extraction [13][14][15], in which target detection is accomplished by embedding a cosine function or embedding a class of clusters in pixels. The other is based on frame-based feature extraction, but this method of embedding clusters has two major disadvantages in the extraction process [16].…”

“…Another feature extraction and positioning method is based on bounding box object detection. [13,14,[17][18][19][20]. [13] addressed two limitations brought up by conventional anchor-based detection: (1) heuristic-guided feature selection; and (2) overlap-based anchor sampling.…”

“…[13] addressed two limitations brought up by conventional anchor-based detection: (1) heuristic-guided feature selection; and (2) overlap-based anchor sampling. Specifically, an anchor-free branch is attached to each level of the feature pyramid, allowing box encoding and decoding in the anchor-free manner at an arbitrary level [14]. Han et al proposed an efficient framework for real-time object tracking which is an end-to-end trained offline Fully Conventional Anchor-Free Siamese network; the network consists of correlation section, implemented by depth-wise cross correlation, and supervised section which has two branches, one for classification and the other for regression.…”