Adaptive Decontamination of the Training Set: A Unified Formulation for Discriminative Visual Tracking

By Martin Danelljan, Gustav Häger, Fahad Shahbaz Khan, Michael Felsberg. Computer Vision and Pattern Recognition, 2016

403 citations recorded by the source corpus.

Tracking-by-detection methods have demonstrated competitive performance in recent years. In these approaches, the tracking model heavily relies on the quality of the training set. Due to the limited amount of labeled training data, additional samples need to be extracted and labeled by the tracker itself. This often leads to the inclusion of corrupted training samples, due to occlusions, misalignments and other perturbations. Existing tracking-by-detection methods either ignore this problem, or employ a separate component for managing the training set. We propose a novel generic approach for alleviating the problem of corrupted training samples in tracking-by-detection frameworks. Our approach dynamically manages the training set by estimating the quality of the samples. Contrary to existing approaches, we propose a unified formulation by minimizing a single loss over both the target appearance model and the sample quality weights. The joint formulation enables corrupted samples to be downweighted while increasing the impact of correct ones. Experiments are performed on three benchmarks: OTB-2015 with 100 videos, VOT-2015 with 60 videos, and Temple-Color with 128 videos. On the OTB-2015, our unified formulation significantly improves the baseline, with a gain of 3:8% in mean overlap precision. Finally, our method achieves state-of-the-art results on all three datasets.

Subfields: Computer Vision and Pattern Recognition, Safety, Risk, Reliability and Quality

View Paper: https://www.semanticscholar.org/paper/992efcca992733f4b170309fbdcd26219f1c0fbb

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Safety, Risk, Reliability and Quality

Citations in the corpus

GOT-10k: A Large High-Diversity Benchmark for Generic Object Tracking in the Wild (2018, methodology): “The same phenomenon can also be observed by comparing DSST and its improved version fDSST, and SRDCF and the improved tracker SRDCFdecon.”
Learning Background-Aware Correlation Filters for Visual Tracking (2017, methodology): “Several recent works addressed the constraint of learning from shifted patches by exploiting training samples whose spatial size is much larger than the trained filters [17, 10, 8, 9].”
Learning Adaptive Discriminative Correlation Filters via Temporal Consistency Preserving Spatial Feature Selection for Robust Visual Object Tracking (2018, methodology): “Furthermore, adaptive decontamination of the training set [57] was proposed to achieve adaptive multi-frame learning in the DCF paradigm, which improved the generalisation performance.”
The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking (2018, background): “These trackers can be generally categorized into 3 classes based on their learning strategy and utilized features: I) correlation ﬁlter (CF) trackers with hand crafted features (KCF [21], Staple-CA [32], and SRDCFdecon [11]); II) CF trackers with deep features (ECO [9], C-COT [12], HDT [36], CF2 [28”
Multi-task Correlation Particle Filter for Robust Object Tracking (2017, methodology): “art tracking methods including MEEM [39], TGPR [12], KCF [16], MUSTer [18], DSST [9], LCT [26], CF2 [25], HDT [29], Staple [2], SRDCF [10], DeepSRDCF [7], SRDCFdecon [8], CNN-SVM [17], and C-COT [11].”
Beyond Correlation Filters: Learning Continuous Convolution Operators for Visual Tracking (2016, methodology): “A further gain in performance is expected by incorporating the unified learning framework [12] to handle corrupted training samples.”
Online Multi-Expert Learning for Visual Tracking (2020, neutral): “We compare the running speed with these two start-of-the-art trackers and the results can be seen in Table III, which shows our running speed is around 14 times faster than CCOT, and around 2 times faster than SRDCFdecon.”
Multi-cue Correlation Filters for Robust Visual Tracking (2018, methodology): “The SRDCF tracker [12] alleviates the boundary effects by penalizing correlation filter coefficients depending on spatial location, and is enhanced by reducing the influence of corrupted samples [10].”
Learning the Model Update for Siamese Trackers (2019, background): “[10, 11] propose to include a subset of historic frames as training samples when calculating the current correlation ﬁlter, which leads to better results than the traditional linear frame-by-frame update.”
Learning Attentions: Residual Attentional Siamese Network for High Performance Online Visual Tracking (2018, methodology): “For example, CF2 [35] and DeepSRDCF [11] concatenated features from different layers of a pretrained CNN such as VGG [40] into correlation filter.”
Learning Target Candidate Association to Keep Track of What Not to Track (2021, background): “[14] propose to learn the tracking model and the training sample weights jointly.”
Learning Dynamic Memory Networks for Object Tracking (2018, methodology): “To further show our tracking accuracy, we also compared with another 8 recent state-of-the art trackers that are not real-time speed, including CREST [26], CSR-DCF [19], MCPF [38], SRDCFdecon [9], SINT [29], SRDCF [6], HDT [23], HCF [20] on OTB-2015.”
VITAL: VIsual Tracking via Adversarial Learning (2018, methodology): “We compare VITAL with 29 trackers from the OTB-2013 benchmark [59] and other 28 stateof-the-art trackers including DSST [10], KCF [22], TGPR [16], MEEM [63], RPT [34], LCT [37], MUSTer [25], HCFT [36], FCNT [57], SRDCF [12], CNN-SVM [24], DeepSRDCF [11], DAT [41], Staple [3], SRDCFdecon [13], CCOT [”
CREST: Convolutional Residual Learning for Visual Tracking (2017, methodology): “We compare our CREST tracker with 29 trackers from the OTB-2013 benchmark [45] and other 21 state-of-the-art trackers including KCF [17], MEEM [48], TGPR [13], DSST [7], RPT [27], MUSTer [20], LCT [31], HCFT [30], FCNT [42], SRDCF [9], CNN-SVM [19], DeepSRDCF [8], DAT [35], Staple [2], SRDCFdecon [1”
Learning Spatial-Temporal Regularized Correlation Filters for Visual Tracking (2018, methodology): “DSST [11] ECO [9] Staple [1] MDNet N [32] TCNN [31] SRDCF [13] BACF [20] SRDCFDecon [10] DeepSRDCF [12] ECO-HC [9] STRCF DeepSTRCF”
Unsupervised Deep Tracking (2019, background): “When integrated with additional improvements such as the adaptive online model update [9, 7], the proposed tracker exhibits state-of-the-art performance.”
AutoTrack: Towards High-Performance Visual Tracking for UAV With Automatic Spatio-Temporal Regularization (2020, methodology): “, SRDCF [8], SAMF [6], CSRDCF [30], SRDCFdecon [24], ARCF-HC [26] are used for comparison.”
ECO: Efficient Convolution Operators for Tracking (2016, methodology): “The recent advancement in DCF based tracking performance is driven by the use of multi-dimensional features [13, 15], robust scale estimation [7, 11], non-linear kernels [20], long-term memory components [28], sophisticated learning models [3, 10] and reducing boundary effects [9, 16].”