Abstract:Most end-to-end discriminative trackers obtain the improvements by a large margin through mining more representative target features for localizing the target, yet it isn't trivial to get comprehensive features as a result of diverse challenges. Moreover, we note that prevalent discriminative trackers merely use a convolution block to process pre-trained residual features to acquire target classification features, which further is input into the appearance model to obtain a target classifier, aiming at obtaining a response map. In view of this problem, we propose a novel feature enhancement module to obtain a richer and comprehensive feature representation, which is end-to-end trainable. Furthermore, discriminative trackers equipped with an online update mechanism demand to refine the classification model with recent samples, which will more or less learn the inaccurate tracking results into the model, thus weakening its discriminative ability. To alleviate this issue, we use a metric learning method to devise a verifier, which is devoted to yielding a similarity score to make piecewise predictions. By comparing the similarity score and piecewise thresholds, this strategy can enhance the robustness of the tracker by adaptively regulating the update of the sample memory and the learning rate. It is worth mentioning that the verifier module has strong portability. We choose the recent SuperDiMP as a baseline and implement comprehensive experimental tests and analyses of our approach on six popular benchmarks, which confirm that the proposed methods perform favorably against state-of-the-art trackers.

Discriminative Multiple Target Tracking

Discriminative Tracking by Metric Learning

Explicitly Modeling the Discriminability for Instance-Aware Visual Object Tracking

Visual Tracking Via Online Discriminative Multiple Instance Metric Learning

Multi-object tracking via discriminative appearance modeling.

Learning to track multiple targets

Effective visual tracking by pairwise metric learning.

Robust Visual Tracking Via Multiple Discriminative Models with Object Proposals

Adaptive Appearance Modeling with Point-to-Set Metric Learning for Visual Tracking

Single and Multiple Object Tracking Using Log-Euclidean Riemannian Subspace and Block-Division Appearance Model.

Description-Discrimination Collaborative Tracking

Enhancing Discriminative Appearance Model for Visual Tracking.

Online Multi-Person Tracking Based on Metric Learning.

Discriminative Metric Learning For Shape Variation Object Tracking

Optimal Metric Evaluation-Based Multicue Inverse Sparse Appearance Model for Object Tracking

Learning Spatio-Temporal Discriminative Model for Affine Subspace Based Visual Object Tracking

Jointly Learning a Multi-class Discriminative Dictionary for Robust Visual Tracking.

Learning Feature Extraction and Classification for Tracking Multiple Objects: A Unified Framework

Learning Deep Appearance Feature for Multi-target Tracking

Visual Tracking Via Joint Discriminative Appearance Learning

Robust Joint Discriminative Feature Learning for Visual Tracking