Visual Object Tracking by Using Ranking Loss

Çevikalp, Hakan; Sarıbaş, Hasan; Benligiray, Burak; Kahvecioglu, Sinem

doi:10.1109/iccvw.2019.00280

Cited by 7 publications

(2 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Therefore, any task requiring to process continuous video streams can benefit from D3D including action/activity recognition with datasets UCF-101 [38], HMDB [26], Kinetics [3]; video object detection task such as ImageNet VID [34]; spatiotemporal action localization task such as Atomic Visual Actions (AVA) dataset [10]; video object tracking (VOT) task such as [24,45]; multi-object tracking (MOT) such as [7]; video person re-identification task such as MARS (Motion Analysis and Re-identification Set) dataset [52]; gait recognition task such as Casia-B dataset [50]; video face recognition such as YouTube Faces [43] and many other tasks. Currently, state-of-the-art architectures either use only spatial content by processing the input frame-by-frame ignoring the temporal content [4,42,2] or utilize offline-trained 3D CNN architectures in a non-dynamic way [20,1]. By utilizing our proposed D3D architecture, all video-based computer vision tasks can incorporate temporal information.…”

Section: Related Workmentioning

confidence: 99%

Dissected 3D CNNs: Temporal Skip Connections for Efficient Online Video Processing

Köpüklü¹,

Hörmann²,

Herzog³

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

Convolutional Neural Networks with 3D kernels (3D CNNs) currently achieve state-of-the-art results in video recognition tasks due to their supremacy in extracting spatiotemporal features within video frames. There have been many successful 3D CNN architectures surpassing the state-of-the-art results successively. However, nearly all of them are designed to operate offline creating several serious handicaps during online operation. Firstly, conventional 3D CNNs are not dynamic since their output features represent the complete input clip instead of the most recent frame in the clip. Secondly, they are not temporal resolutionpreserving due to their inherent temporal downsampling. Lastly, 3D CNNs are constrained to be used with fixed temporal input size limiting their flexibility. In order to address these drawbacks, we propose dissected 3D CNNs, where the intermediate volumes of the network are dissected and propagated over depth (time) dimension for future calculations, substantially reducing the number of computations at online operation. For action classification, the dissected version of ResNet models performs 74-90% fewer computations at online operation while achieving ∼5% better classification accuracy on the Kinetics-600 dataset than conventional 3D ResNet models. Moreover, the advantages of dissected 3D CNNs are demonstrated by deploying our approach onto several vision tasks, which consistently improved the performance.

show abstract

Section: Related Workmentioning

confidence: 99%

Dissected 3D CNNs: Temporal Skip Connections for Efficient Online Video Processing

Köpüklü¹,

Hörmann²,

Herzog³

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…A fast version of MDNet tracker, called Real-Time MDNet, [32] used Fast R-CNN [20] approach to accelerate the slow feature extraction stage of the MDNet method. Cevikalp et al [7] proposed a deep neural network tracker using ranking loss which enforces the network to return better bounding boxes framing the target object. For the same purpose, both [11,3] used novel deep neural tracking architectures that utilize IoU-Net [30] whose goal is to estimate and increase the Intersection over Union (IoU) overlap between the target an estimated bounding box to improve the accuracy.…”

Section: Related Workmentioning

confidence: 99%

TRAT: Tracking by Attention Using Spatio-Temporal Features

Saribas,

Cevikalp,

Köpüklü

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

Robust object tracking requires knowledge of tracked objects' appearance, motion and their evolution over time. Although motion provides distinctive and complementary information especially for fast moving objects, most of the recent tracking architectures primarily focus on the objects' appearance information. In this paper, we propose a twostream deep neural network tracker that uses both spatial and temporal features. Our architecture is developed over ATOM tracker and contains two backbones: (i) 2D-CNN network to capture appearance features and (ii) 3D-CNN network to capture motion features. The features returned by the two networks are then fused with attention based Feature Aggregation Module (FAM). Since the whole architecture is unified, it can be trained end-to-end. The experimental results show that the proposed tracker TRAT (TRacking by ATtention) achieves state-of-the-art performance on most of the benchmarks and it significantly outperforms the baseline ATOM tracker. The source code and models are available at https://github.com/Hasan4825/TRAT.

show abstract