Joint Classification and Regression for Visual Tracking with Fully Convolutional Siamese Networks

Visual tracking of generic objects is one of the fundamental but challenging problems in computer vision. Here, we propose a novel fully convolutional Siamese network to solve visual tracking by directly predicting the target bounding box in an end-to-end manner. We first reformulate the visual tracking task as two subproblems: a classification problem for pixel category prediction and a regression task for object status estimation at this pixel. With this decomposition, we design a simple yet effective Siamese architecture based classification and regression framework, termed SiamCAR, which consists of two subnetworks: a Siamese subnetwork for feature extraction and a classification-regression subnetwork for direct bounding box prediction. Since the proposed framework is both proposal- and anchor-free, SiamCAR can avoid the tedious hyper-parameter tuning of anchors, considerably simplifying the training. To demonstrate that a much simpler tracking framework can achieve superior tracking results, we conduct extensive experiments and comparisons with state-of-the-art trackers on a few challenging benchmarks. Without bells and whistles, SiamCAR achieves leading performance with a real-time speed. Furthermore, the ablation study validates that the proposed framework is effective with various backbone networks, and can benefit from deeper networks. Code is available at https://github.com/ohhhyeahhh/SiamCAR.

Robust Data Association Using Fusion of Data-Driven and Engineered Features for Real-Time Pedestrian Tracking in Thermal Images

Object tracking is an essential problem in computer vision that has been extensively researched for decades. Tracking objects in thermal images is particularly difficult because of the lack of color information, low image resolution, or high similarity between objects of the same class. One of the main challenges in multi-object tracking, also referred to as the data association problem, is finding the correct correspondences between measurements and tracks and adapting the object appearance changes over time. We addressed this challenge of data association for thermal images by proposing three contributions. The first contribution consisted of the creation of a data-driven appearance score using five Siamese Networks, which operate on the image detection and on parts of it. Secondly, we engineered an original edge-based descriptor that improves the data association process. Lastly, we proposed a dataset consisting of pedestrian instances that were recorded in different scenarios and are used for training the Siamese Networks. The data-driven part of the data association score offers robustness, while feature engineering offers adaptability to unknown scenarios and their combination leads to a more powerful tracking solution. Our approach had a running time of 25 ms and achieved an average precision of 86.2% on publicly available benchmarks, containing real-world scenarios, as shown in the evaluation section.

Siamese anchor-free object tracking with multiscale spatial attentions

Recently, object trackers based on Siamese networks have attracted considerable attentions due to their remarkable tracking performance and widespread application. Especially, the anchor-based methods exploit the region proposal subnetwork to get accurate prediction of a target and make great performance improvement. However, those trackers cannot capture the spatial information very well and the pre-defined anchors will hinder robustness. To solve these problems, we propose a Siamese-based anchor-free object tracking algorithm with multiscale spatial attentions in this paper. Firstly, we take ResNet-50 as the backbone network to generate multiscale features of both template patch and search regions. Secondly, we propose the spatial attention extraction (SAE) block to capture the spatial information among all positions in the template and search region feature maps. Thirdly, we put these features into the SAE block to get the multiscale spatial attentions. Finally, an anchor-free classification and regression subnetwork is used for predicting the location of the target. Unlike anchor-based methods, our tracker directly predicts the target position without predefined parameters. Extensive experiments with state-of-the-art trackers are carried out on four challenging visual object tracking benchmarks: OTB100, UAV123, VOT2016 and GOT-10k. Those experimental results confirm the effectiveness of our proposed tracker.

Learning Geometry Information of Target for Visual Object Tracking with Siamese Networks

Recently, Siamese architecture has been widely used in the field of visual tracking, and has achieved great success. Most Siamese network based trackers aggregate the target information of two branches by cross-correlation. However, since the location of the sampling points in the search feature area is pre-fixed in cross-correlation operation, these trackers suffer from either background noise influence or missing foreground information. Moreover, the cross-correlation between the template and the search area neglects the geometry information of the target. In this paper, we propose a Siamese deformable cross-correlation network to model the geometric structure of target and improve the performance of visual tracking. We propose to learn an offset field end-to-end in cross-correlation. With the guidance of the offset field, the sampling in the search image area can adapt to the deformation of the target, and realize the modeling of the geometric structure of the target. We further propose an online classification sub-network to model the variation of target appearance and enhance the robustness of the tracker. Extensive experiments are conducted on four challenging benchmarks, including OTB2015, VOT2018, VOT2019 and UAV123. The results demonstrate that our tracker achieves state-of-the-art performance.