1. |
- Barbalau, Antonio, et al.
(author)
-
SSMTL plus plus : Revisiting self-supervised multi-task learning for video anomaly detection
- 2023
-
In: Computer Vision and Image Understanding. - : ACADEMIC PRESS INC ELSEVIER SCIENCE. - 1077-3142 .- 1090-235X. ; 229
-
Journal article (peer-reviewed)abstract
- A self-supervised multi-task learning (SSMTL) framework for video anomaly detection was recently introduced in literature. Due to its highly accurate results, the method attracted the attention of many researchers. In this work, we revisit the self-supervised multi-task learning framework, proposing several updates to the original method. First, we study various detection methods, e.g. based on detecting high-motion regions using optical flow or background subtraction, since we believe the currently used pre-trained YOLOv3 is suboptimal, e.g. objects in motion or objects from unknown classes are never detected. Second, we modernize the 3D convolutional backbone by introducing multi-head self-attention modules, inspired by the recent success of vision transformers. As such, we alternatively introduce both 2D and 3D convolutional vision transformer (CvT) blocks. Third, in our attempt to further improve the model, we study additional self-supervised learning tasks, such as predicting segmentation maps through knowledge distillation, solving jigsaw puzzles, estimating body pose through knowledge distillation, predicting masked regions (inpainting), and adversarial learning with pseudo-anomalies. We conduct experiments to assess the performance impact of the introduced changes. Upon finding more promising configurations of the framework, dubbed SSMTL++v1 and SSMTL++v2, we extend our preliminary experiments to more data sets, demonstrating that our performance gains are consistent across all data sets. In most cases, our results on Avenue, ShanghaiTech and UBnormal raise the state-of-the-art performance bar to a new level.
|
|
2. |
|
|
3. |
- Björkman, Mårten, 1970-, et al.
(author)
-
Detecting, segmenting and tracking unknown objects using multi-label MRF inference
- 2014
-
In: Computer Vision and Image Understanding. - : Elsevier. - 1077-3142 .- 1090-235X. ; 118, s. 111-127
-
Journal article (peer-reviewed)abstract
- This article presents a unified framework for detecting, segmenting and tracking unknown objects in everyday scenes, allowing for inspection of object hypotheses during interaction over time. A heterogeneous scene representation is proposed, with background regions modeled as a combinations of planar surfaces and uniform clutter, and foreground objects as 3D ellipsoids. Recent energy minimization methods based on loopy belief propagation, tree-reweighted message passing and graph cuts are studied for the purpose of multi-object segmentation and benchmarked in terms of segmentation quality, as well as computational speed and how easily methods can be adapted for parallel processing. One conclusion is that the choice of energy minimization method is less important than the way scenes are modeled. Proximities are more valuable for segmentation than similarity in colors, while the benefit of 3D information is limited. It is also shown through practical experiments that, with implementations on GPUs, multi-object segmentation and tracking using state-of-art MRF inference methods is feasible, despite the computational costs typically associated with such methods.
|
|
4. |
- Bretzner, Lars, et al.
(author)
-
Feature Tracking with Automatic Selection of Spatial Scales
- 1998
-
In: Computer Vision and Image Understanding. - : Elsevier BV. - 1077-3142 .- 1090-235X. ; 71:3, s. 385-393
-
Journal article (peer-reviewed)abstract
- When observing a dynamic world, the size of image structures may vary over time. This article emphasizes the need for including explicit mechanisms for automatic scale selection in feature tracking algorithms in order to: (i) adapt the local scale of processing to the local image structure, and (ii) adapt to the size variations that may occur over time. The problems of corner detection and blob detection are treated in detail, and a combined framework for feature tracking is presented. The integrated tracking algorithm overcomes some of the inherent limitations of exposing fixed-scale tracking methods to image sequences in which the size variations are large. It is also shown how the stability over time of scale descriptors can be used as a part of a multi-cue similarity measure for matching. Experiments on real-world sequences are presented showing the performance of the algorithm when applied to (individual) tracking of corners and blobs.
|
|
5. |
|
|
6. |
- Croitoru, Florinel-Alin, et al.
(author)
-
Lightning fast video anomaly detection via multi-scale adversarial distillation
- 2024
-
In: Computer Vision and Image Understanding. - : ACADEMIC PRESS INC ELSEVIER SCIENCE. - 1077-3142 .- 1090-235X. ; 247
-
Journal article (peer-reviewed)abstract
- We propose a very fast frame-level model for anomaly detection in video, which learns to detect anomalies by distilling knowledge from multiple highly accurate object-level teacher models. To improve the fidelity of our student, we distill the low-resolution anomaly maps of the teachers by jointly applying standard and adversarial distillation, introducing an adversarial discriminator for each teacher to distinguish between target and generated anomaly maps. We conduct experiments on three benchmarks (Avenue, ShanghaiTech, UCSD Ped2), showing that our method is over 7 times faster than the fastest competing method, and between 28 and 62 times faster than object-centric models, while obtaining comparable results to recent methods. Our evaluation also indicates that our model achieves the best trade-off between speed and accuracy, due to its previously unheard-of speed of 1480 FPS. In addition, we carry out a comprehensive ablation study to justify our architectural design choices. Our code is freely available at: https://github.com/ristea/fast-aed.
|
|
7. |
- Hoang, Tuan, et al.
(author)
-
Simultaneous compression and quantization: A joint approach for efficient unsupervised hashing
- 2020
-
In: Computer Vision and Image Understanding. - : Elsevier BV. - 1077-3142 .- 1090-235X. ; 191
-
Journal article (peer-reviewed)abstract
- For unsupervised data-dependent hashing, the two most important requirements are to preserve similarity in the low-dimensional feature space and to minimize the binary quantization loss. A well-established hashing approach is Iterative Quantization (ITQ), which addresses these two requirements in separate steps. In this paper, we revisit the ITQ approach and propose novel formulations and algorithms to the problem. Specifically, we propose a novel approach, named Simultaneous Compression and Quantization (SCQ), to jointly learn to compress (reduce dimensionality) and binarize input data in a single formulation under strict orthogonal constraint. With this approach, we introduce a loss function and its relaxed version, termed Orthonormal Encoder (OnE) and Orthogonal Encoder (OgE) respectively, which involve challenging binary and orthogonal constraints. We propose to attack the optimization using novel algorithms based on recent advance in cyclic coordinate descent approach. Comprehensive experiments on unsupervised image retrieval demonstrate that our proposed methods consistently outperform other state-of-the-art hashing methods. Notably, our proposed methods outperform recent deep neural networks and GAN based hashing in accuracy, while being very computationally-efficient.
|
|
8. |
- Ismaeil, Kassem Al, et al.
(author)
-
Enhancement of dynamic depth scenes by upsampling for precise super-resolution (UP-SR)
- 2016
-
In: Computer Vision and Image Understanding. - : Academic Press Inc Elsevier. - 1077-3142 .- 1090-235X. ; 147, s. 38-49
-
Journal article (peer-reviewed)abstract
- Multi-frame super-resolution is the process of recovering a high resolution image or video from a set of captured low resolution images. Super-resolution approaches have been largely explored in 2-D imaging. However, their extension to depth videos is not straightforward due to the textureless nature of depth data, and to their high frequency contents coupled with fast motion artifacts. Recently, few attempts have been introduced where only the super-resolution of static depth scenes has been addressed. In this work, we propose to enhance the resolution of dynamic depth videos with non-rigidly moving objects. The proposed approach is based on a new data model that uses densely upsampled, and cumulatively registered versions of the observed low resolution depth frames. We show the impact of upsampling in increasing the sub-pixel accuracy and reducing the rounding error of the motion vectors. Furthermore, with the proposed cumulative motion estimation, a high registration accuracy is achieved between non-successive upsampled frames with relative large motions. A statistical performance analysis is derived in terms of mean square error explaining the effect of the number of observed frames and the effect of the super-resolution factor at a given noise level. We evaluate the accuracy of the proposed algorithm theoretically and experimentally as function of the SR factor, and the level of noise contamination. Experimental results on both real and synthetic data show the effectiveness of the proposed algorithm on dynamic depth videos as compared to state-of-art methods.
|
|
9. |
- Julià, Carme, et al.
(author)
-
Shape-based image segmentation through photometric stereo
- 2011
-
In: Computer Vision and Image Understanding. - : Elsevier BV. - 1077-3142 .- 1090-235X. ; 115:1, s. 91-104
-
Journal article (peer-reviewed)abstract
- This paper describes a new algorithm for segmenting 2D images by taking into account 3D shape information. The proposed approach consists of two stages. In the first stage, the 3D surface normals of the objects present in the scene are estimated through robust photometric stereo. Then, the image is segmented by grouping its pixels according to their estimated normals through graph-based clustering. One of the advantages of the proposed approach is that, although the segmentation is based on the 3D shape of the objects, the photometric stereo stage used to estimate the 3D normals only requires a set of 2D images. This paper provides an extensive validation of the proposed approach by comparing it with several image segmentation algorithms. Particularly, it is compared with both appearance-based image segmentation algorithms and shape-based ones. Experimental results confirm that the latter are more suitable when the objective is to segment the objects or surfaces present in the scene. Moreover, results show that the proposed approach yields the best image segmentation in most of the cases.
|
|
10. |
- Khan, Zulfiqar Hasan, 1976, et al.
(author)
-
Online Domain-Shift Learning and Object Tracking based on Nonlinear Dynamic Models and Particle Filters on Riemannian Manifolds
- 2014
-
In: Computer Vision and Image Understanding. - : Elsevier BV. - 1077-3142 .- 1090-235X. ; 125, s. 97-114
-
Journal article (peer-reviewed)abstract
- This paper proposes a novel online domain-shift appearance learning and object tracking scheme on a Riemannian manifold for visual and infrared videos, especially for video scenarios containing large deformable objects with fast out-of-plane pose changes that could be accompanied by partial occlusions. Although Riemannian manifolds and covariance descriptors are promising for visual object tracking, the use of Riemannian mean from a window of observations, spatially insensitive covariance descriptors, fast significant out-of-plane (non-planar) pose changes, and long-term partial occlusions of large-size deformable objects in video limits the performance of such trackers. The proposed method tackles these issues with the following main contributions: (a) Proposing a Bayesian formulation on Riemannian manifolds by using particle filters on the manifold and using appearance particles in each time instant for computing the Riemannian mean, rather than using a window of observations. (b) Proposing a nonlinear dynamic model for online domain-shift learning on the manifold, where the model includes both manifold object appearance and its velocity. (c) Introducing a criterion-based partial occlusion handling approach in online learning. (d) Tracking object bounding box by using affine parametric shape modeling with manifold appearance embedded. (e) Incorporating spatial, frequency and orientation information in the covariance descriptor by extracting Gabor features in a partitioned bounding box. (f) Effectively applying to both visual-band videos and thermal-infrared videos. To realize the proposed tracker, two particle filters are employed: one is applied on the Riemannian manifold for generating candidate appearance particles and another is on vector space for generating candidate box particles. Further, tracking and online learning are performed in alternation to mitigate the tracking drift. Experiments on both visual and infrared videos have shown robust tracking performance of the proposed scheme. Comparisons and evaluations with ten existing state-of-art trackers provide further support to the proposed scheme.
|
|