Correlation-Aware Deep Tracking

Xie, Fei; Wang, Qianqian; Wang, Guangting; Cao, Yue; Yang, Wankou; Zeng, Wenjun

doi:10.1109/cvpr52688.2022.00855

Cited by 115 publications

(22 citation statements)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For example, inspired by Transformers, TransT [6] uses attentionbased feature fusion to combine features of the object template and search image. More recently, several works utilize Transformers as direct predictors to achieve a new state of the art, such as STARK [71], ToMP [47] and SBT [70]. These models tokenize frame features from a ResNet [25] encoder, and use a Transformer to predict the bounding box and object presence score with the feature tokens.…”

Section: Single Object Tracking Methodologiesmentioning

confidence: 99%

EgoTracks: A Long-term Egocentric Visual Object Tracking Dataset

Tang¹,

Liang²,

Grauman³

et al. 2023

Preprint

View full text Add to dashboard Cite

Figure 1. A sample video from the proposed EgoTracks dataset, with yellow segments of the clip marking when the object (blowtorch) is visible. Note the frequent disappearances and reappearances of the object over an 8 minute video, with lengthy absences, necessitating re-detection to track accurately without false positives. The egocentric nature of the video includes the camera-wearer interacting with the object (Occurrence 2), resulting in significant hand occlusions and dramatic changes in pose.

show abstract

Section: Single Object Tracking Methodologiesmentioning

confidence: 99%

EgoTracks: A Long-term Egocentric Visual Object Tracking Dataset

Tang¹,

Liang²,

Grauman³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…In recent years, some 2D one-stream frameworks [33], [34] have been proposed to perform their unique character in 2D object tracking. Ye et al [33] released some Siamesebased framework's problem: the extracted features lack the awareness of the target and have limited target-background discriminability.…”

Section: D One-stream Trackingmentioning

confidence: 99%

“…To tackle these issues, Ye et alintroduced a one-stream tracking framework which can bridge the templatesearch image pairs with bidirectional information flows. Single Branch Transformer (SBT) [34] suppresses non-target features and obtain instance-varying features by extensively matching the features of the two images through one-stream backbone. To sum up, one-stream frameworks possess the ability to extract more discriminative features than ordinary Siamese frameworks.…”

Section: D One-stream Trackingmentioning

confidence: 99%

OST: Efficient One-stream Network for 3D Single Object Tracking in Point Clouds

Zhao¹,

Han²,

Tian³

et al. 2022

Preprint

View full text Add to dashboard Cite

Although recent Siamese network-based trackers have achieved impressive perceptual accuracy for single object tracking in LiDAR point clouds, they advance with some heavy correlation operations on relation modeling and overlook the inherent merit of arbitrariness compared to multiple object tracking. In this work, we propose a radically novel one-stream network with the strength of the Transformer encoding, which avoids the correlation operations occurring in previous Siamese network, thus considerably reducing the computational effort. In particular, the proposed method mainly consists of a Templateaware Transformer Module (TTM) and a Multi-scale Feature Aggregation (MFA) module capable of fusing spatial and semantic information. The TTM stitches the specified template and the search region together and leverages an attention mechanism to establish the information flow, breaking the previous pattern of independent extraction-and-correlation. As a result, this module makes it possible to directly generate template-aware features that are suitable for the arbitrary and continuously changing nature of the target, enabling the model to deal with unseen categories. In addition, the MFA is proposed to make spatial and semantic information complementary to each other, which is characterized by reverse directional feature propagation that aggregates information from shallow to deep layers. Extensive experiments on KITTI and nuScenes demonstrate that our method has achieved considerable performance not only for classspecific tracking but also for class-agnostic tracking with less computation and higher efficiency.

show abstract

“…Xie et al [36] found that the features extracted by Siamese-like networks cannot completely distinguish between the tracked target and the distractor objects. Therefore, by deeply embedding cross-image feature correlation in multiple layers of the feature network, a new target-dependent feature network is proposed, and it allows the features of the search and template images to be deeply fused for tracking.…”

Section: Related Workmentioning

confidence: 99%

Object Tracking Based on a Time-Varying Spatio-Temporal Regularized Correlation Filter With Aberrance Repression

et al. 2022

View full text Add to dashboard Cite

When used for object tracking, the discriminative correlation filter (DCF) is effective, but its performance is often burdened by undesirable boundary effects. Meanwhile, when there is too much background information in training samples of the DCF, it will be easier to learn the area deviating from the tracking object. Further, factors such as illumination variation, partial/full occlusion of the object, and variations in the object appearance, render the response map aberrance of the correlation filter (CF) more prone to occur. To overcome the aforementioned problems, an object tracking model based on a time-varying Spatio-temporal regularized correlation filter with aberrance repression is proposed in this paper. Firstly, by adding a regularized term to the objective function of the traditional CFs to limit the change rate of the response map generated in the object detection phase, the proposed tracker can obviously repress the aberrance of the response maps; secondly, by adjusting the filter to the object regions suitable for tracking with high confidence scores with a time-varying spatial reliability map, the proposed tracker effectively overcomes the adverse effects caused by the boundary effect; and finally, by introducing a temporal regularized term, the proposed tracker also has superior tracking ability for the partial occluded objects and those with large appearance variations. Significant experiments on the OTB100, VOT2016, TC128, and UAV 123 datasets have revealed that although the proposed tracker only uses the histogram of directional gradient (HOG) and color name (CN) features, the performance thereof outperformed many state-ofthe-art trackers based on DCF and deep-based frameworks in terms of tracking accuracy, tracking success rate, and A-R rank, etc.

show abstract

Correlation-Aware Deep Tracking

Cited by 115 publications

References 39 publications

EgoTracks: A Long-term Egocentric Visual Object Tracking Dataset

EgoTracks: A Long-term Egocentric Visual Object Tracking Dataset

OST: Efficient One-stream Network for 3D Single Object Tracking in Point Clouds

Object Tracking Based on a Time-Varying Spatio-Temporal Regularized Correlation Filter With Aberrance Repression

Contact Info

Product

Resources

About