Unsupervised Deep Tracking

By Ning Wang, Yibing Song, Chao Ma, Wen-gang Zhou, Wei David Liu, Houqiang Li. Computer Vision and Pattern Recognition, 2019

361 citations recorded by the source corpus.

We propose an unsupervised visual tracking method in this paper. Different from existing approaches using extensive annotated data for supervised learning, our CNN model is trained on large-scale unlabeled videos in an unsupervised manner. Our motivation is that a robust tracker should be effective in both the forward and backward predictions (i.e., the tracker can forward localize the target object in successive frames and backtrace to its initial position in the first frame). We build our framework on a Siamese correlation filter network, which is trained using unlabeled raw videos. Meanwhile, we propose a multiple-frame validation method and a cost-sensitive loss to facilitate unsupervised learning. Without bells and whistles, the proposed unsupervised tracker achieves the baseline accuracy of fully supervised trackers, which require complete and accurate labels during training. Furthermore, unsupervised framework exhibits a potential in leveraging unlabeled or weakly labeled data to further improve the tracking accuracy.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/8b6555c46a2d02e713dbd339e7ac7230ace5193f

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

Citations in the corpus

Space-Time Correspondence as a Contrastive Random Walk (2020, methodology): “In lieu of ground truth labels, we use the idea of cycle-consistency [99, 88, 84], by turning training videos into palindromes , where the ﬁrst half is repeated backwards.”
Learning Video Object Segmentation From Unlabeled Videos (2020, methodology): “To capture this property, we apply a forward-backward patch tracking mechanism [57] which learns φ by tracking a sampled patch forwards in a few successive frames and then backwards until the start frame, and penalizing the distance between the initial and final backwards tracked positions of that p”
AutoTrack: Towards High-Performance Visual Tracking for UAV With Automatic Spatio-Temporal Regularization (2020, methodology): “Compared to deep trackers including ASRCF [31], TADT [51], SiameseFC [15], DSiam [56], MCCT [13], ADNet [16], CFNet [52], DeepSTRCF [14], UDT+ [53], HCF [18], C-COT [17], ECO [20], IBCCF [54], MCPF [19] and CREST [35], AutoTrack with a single CPU exhibits the”
Towards More Flexible and Accurate Object Tracking with Natural Language: Algorithms and Benchmark (2021, neutral): “285 035 UDT [44] CVPR-2019 CNN-5 7 3 Local 70@GPU 0.”
Ocean: Object-aware Anchor-free Tracking (2020, methodology): “There are 9 anchor-based Siamese framework based methods (SiamFC [1], GradNet [24], DSiam [11], MemDTC [46], SiamRPN [22], C-RPN [9], SiamMASK [43], SiamRPN++ [21] and SiamRCNN [39]), 8 discriminative correlation filter based methods (CFNet [38], ECO [5], STRCF [23], LADCF [45], UDT [42], STN [37], ”
Deep Learning for Visual Tracking: A Comprehensive Survey (2019, methodology): “The main contributions of the SNN-based methods are: i) exploiting multi-stage Siamese framework to stimulate hard negative sampling [157], ii) adopting sampling heuristic such as fixed foregroundto-background ratio [157] or sampling strategies such as random sampling [111] or flow-guided sampling [”
LaSOT: A High-quality Large-scale Single Object Tracking Benchmark (2020, background): “Owing to balanced efficiency and accuracy, deep Siamese tracking has been extended by many later works [27, 36, 49, 51, 78, 80, 95].”
Transformer Meets Tracker: Exploiting Temporal Context for Robust Visual Tracking (2021, background): “The recent improvements upon Siamese trackers include attention mechanism [54], reinforcement learning [22, 51], target-aware model fine-tuning [30], unsupervised training [50, 52], sophisticated backbone networks [27, 61], cascaded frameworks [14, 49], and model update mechanisms [16, 17, 58, 60].”