Abstract:Current Siamese trackers have paid more attention to Transformer-based structures for their extraordinary improvements in accuracy through extensive information fusion and cross-attention enhancement. However, the further elevation of traditional Siamese trackers' performance is lagged by the low robustness of the interference from similar distractors. The representation ability and the discrimination against similar distractors are always incompatible. Even though the representation ability of recent Transformer-based trackers is broadly enhanced, it still causes a high response to similar distractors owing to the similarity matching mechanism of the Siamese structure. To tackle the above problems, we propose a Tracking-in-Tracking (an outer tracker with an inner tracker) pipeline (TiT) consisting of an antecedent tracking stage and a refining tracking stage. Instead of just capturing a single candidate matched with the template, perceiving all potential candidates can provide proper information on possible similar distractors. Based on this insight, a Transformer-based outer tracker is constructed to recognize all candidates in the antecedent tracking stage. Subsequently, in the refining tracking stage, an inner tracker is applied to further realize accurate object identification from all selected candidates with a designed bilateral feedback mechanism (BFM) and peak distilling module (PDM). Therefore, the Transformer-based outer tracker and Motion-estimated inner tracker can supervise each other to achieve robust tracking performance without further aggravating model complexity and memory burden. Extensive experiments have demonstrated that our TiT can serve as a unified framework to discriminate similar interference and perform state-of-the-art (SOTA) performance in mainstream benchmarks.

An Approach to the Correspondence Problem in the 1-D Optical Transducers Tracking System

A Systemic Model for Simultaneously Measuring Three dimensional Rotation Angles of the Linear Guide Track [J]

Linear CCD Based Optical Tracking Using Stereo Correspondence Algorithm

A Simple 3d Face Tracking Method Based On Depth Information

An Effective Approach to 3D Deformable Surface Tracking

Positional Accuracy and Transmitter Orientation of the 3D Electromagnetic Tracking System

2D Tracking Algorithm Performance Testing Based on Three- Dimensional Visual Platform

A New Tracking System for Three Magnetic Objectives

Length-measured method for virtual coordinate measurement by laser tracking system

A Combined Optical Positioning Method Based on 1-D and 2-D Cameras

Exploring Multi-Target Tracking Method Using Spatially Multiplexed Superposition Imaging Technique

Tracking in tracking: An efficient method to solve the tracking distortion

A Novel Marker Tracking Method Based on Extended Kalman Filter for Multi-Camera Optical Tracking Systems

Position And Attitude Measurement Of Moving Target Using Laser Tracking System With Multiple Measuring Stations

A Ball-Shaped Target Development and Pose Estimation Strategy for a Tracking-Based Scanning System

Viewpoint placement optimization for binocular three-dimensional trajectory measurement of translating and rotating particle swarms

An Improved Real-Time Visual Tracking Method for Space Non-Cooperative Target

Collaborative Target Tracking in Elliptic Coordinates: a Binocular Coordination Approach

Ccd-Based System For Three-Dimensional Real-Time Positioning

Computer-Vision-Based Vibration Tracking Using a Digital Camera: A Sparse-Optical-Flow-Based Target Tracking Method

An Electromagnetic Tracking Method Using Rotating Orthogonal Coils