ATOM: Accurate Tracking by Overlap Maximization

By Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, Michael Felsberg. Computer Vision and Pattern Recognition, 2018

1,321 citations recorded by the source corpus.

While recent years have witnessed astonishing improvements in visual tracking robustness, the advancements in tracking accuracy have been limited. As the focus has been directed towards the development of powerful classifiers, the problem of accurate target state estimation has been largely overlooked. In fact, most trackers resort to a simple multi-scale search in order to estimate the target bounding box. We argue that this approach is fundamentally limited since target estimation is a complex task, requiring high-level knowledge about the object. We address this problem by proposing a novel tracking architecture, consisting of dedicated target estimation and classification components. High level knowledge is incorporated into the target estimation through extensive offline learning. Our target estimation component is trained to predict the overlap between the target object and an estimated bounding box. By carefully integrating target-specific information, our approach achieves previously unseen bounding box accuracy. We further introduce a classification component that is trained online to guarantee high discriminative power in the presence of distractors. Our final tracking framework sets a new state-of-the-art on five challenging benchmarks. On the new large-scale TrackingNet dataset, our tracker ATOM achieves a relative gain of 15% over the previous best approach, while running at over 30 FPS. Code and models are available at https://github.com/visionml/pytracking.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/d74169a8fd2f90a06480d1d583d0ae5e980ea951

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

Backbone is All Your Need: A Simplified Architecture for Visual Object Tracking (2022, background): “The compared trackers include SiamFC [2](ECCVW16), ATOM [11] (CVPR19), SiamRPN++ [42](CVPR19), DiMP [3](ICCV19), SiamFC++ [42] (AAAI20), Ocean [49](ECCV20), SiamBAN [9](CVPR20), SiamAtt [45] (CVPR 20), TransT [38](CVPR21), TrDiMP [39](CVPR21), KeepTrack [27](ICCV21), STARKS [43](ICCV21), AutoMatch [”
Probabilistic Regression for Visual Tracking (2020, methodology): “The baseline [3, 9], employing L2 loss to predict IoU, achieves 63.”
Transformer Meets Tracker: Exploiting Temporal Context for Robust Visual Tracking (2021, methodology): “Similar to the previous works [7, 3, 9, 2], we utilize the training splits of LaSOT [13], TrackingNet [38], GOT-10k [23], and COCO [31] for offline training.”
Siamese Box Adaptive Network for Visual Tracking (2020, methodology): “MDNet [28] ECO [6] C-COT [8] UPDT [3] ATOM [5] DiMP [2] Ours”
Towards More Flexible and Accurate Object Tracking with Natural Language: Algorithms and Benchmark (2021, methodology): “Tracking by Bounding Box The standard trackers begin their tracking procedure based on an initialized BBox in the first frame, including classification based [19,26,44,45], Siamese network based [9, 10, 30, 53, 59], correlation filter based [11, 12, 22, 40], and regression-based [21].”
STMTrack: Template-free Visual Tracking with Space-time Memory Networks (2021, neutral): “Ours DiMP-50 [2] ATOM [10] SiamBAN [6] SiamCAR [15] SiamRPN++ [23] UPDT [4]”
Energy-Based Models for Deep Probabilistic Regression (2020, methodology): “The other approach is to treat the confidence prediction as a direct regression problem itself by applying standard regression losses, such as L(2) [50, 8, 55] or the Huber loss [23].”
Learning Target Candidate Association to Keep Track of What Not to Track (2021, methodology): “Many online trackers [12, 3, 13] employ a memory to store previous predictions to fine-tune the tracker.”
LaSOT: A High-quality Large-scale Single Object Tracking Benchmark (2020, methodology): “To deal with scale variation, the methods of [15] introduce the intersectionover-union (IoU) network for tracking and achieve promising results.”
Deformable Siamese Attention Networks for Visual Object Tracking (2020, methodology): “An alternative solution is to extend existing on-line discriminative framework with deep networks for end-to-end learning, e.g., ATOM [6] and DiMP [2].”
LasHeR: A Large-Scale High-Diversity Benchmark for RGBT Tracking (2021, methodology): “From the results, we can see that DMCNet achieves the best precision score of 0.557 and normalized precision score of 0.496, while mfDiMP combines the IoU-Net based architecture from ATOM [46], achieving the top success score of 0.399.”
SiamFC++: Towards Robust and Accurate Visual Tracking with Target Estimation Guidelines (2019, methodology): “An estimation quality score independent of classification should be used, as in many previous pieces of research about both object detection and tracking (Jiang et al. 2018; Tian et al. 2019; Danelljan et al. 2019).”
Know Your Surroundings: Exploiting Scene Information for Object Tracking (2020, methodology): “sing the circular shifts of the input image as training data. Another approach is to train or ne-tune a few layers of a deep neural network in the rst frame to perform target-background classication [37,8,3,40]. MDNet [37] ne-tunes three fully-connected layers online, while DiMP [3] employs a meta-l”
Learning the Model Update for Siamese Trackers (2019, methodology): “[8] [7] [40] [32] Linear UpdateNet Linear UpdateNet”
Segment as Points for Efficient Online Multi-Object Tracking and Segmentation (2020, background): “ion. Detection based MOT approaches rst detect objects of interests and then link objects into trajectories via data association. The data association can be accomplished on either the 2D image plane [4,6,7,14,33,42,38] or the 3D world space [1,8,10,20,25,39]. ATOM [7] introduces a novel tracking ar”
Ocean: Object-aware Anchor-free Tracking (2020, methodology): “We refer readers to [2,4] for more details.”