86,876 research outputs found
Layered representation of a video shot with mosaicing
No full textThis paper presents a motion segmentation method useful for representing efficiently a video shot as a static mosaic of the background plus sequences of the objects moving in the foreground. This generates an MPEG-4 compliant, layered representation useful for video coding, editing and indexing. First, a mosaic of the static background is computed by estimating the dominant motion of the scene. This is achieved by tracking features over the video sequence and using a robust technique that discards features attached to the moving objects. The moving objects get removed in the final mosaic by computing the median of the grey levels. Then, segmentation is obtained by taking the pixelwise difference between each frame of the original sequence and the mosaic of the background. To discriminate between the moving object and noise, temporal coherence is exploited by tracking the object in the binarised difference image sequence. The automatic computation of the mosaic and the segmentation procedure are illustrated with real sequences experiments. Examples of coding and content-based manipulation are also shown.</p
FasterVideo: Efficient Online Joint Object Detection and Tracking
No full textObject detection and tracking in videos represent essential and computationally demanding building blocks for current and future visual perception systems. In order to reduce the efficiency gap between available methods and computational requirements of real-world applications, we propose to re-think one of the most successful methods for image object detection, Faster R-CNN, and extend it to the video domain. Specifically, we extend the detection framework to learn instance-level embeddings which prove beneficial for data association and re-identification purposes. Focusing on the computational aspects of detection and tracking, our proposed method reaches a very high computational efficiency necessary for relevant applications, while still managing to compete with recent and state-of-the-art methods as shown in the experiments we conduct on standard object tracking benchmarks (Code available at https://github.com/Malga-Vision/fastervideo )
Adversarial feature refinement for cross-view action recognition
No full textApparent motion information of an action may vary dramatically from one view to another, making transfer of knowledge across views a core challenge of action recognition. Recent times have seen the use of large scale datasets to compensate for this lack in generalization, and in fact most state-of-the-art methods today require large amounts of training data and have high computational cost while training. We propose a novel technique leveraging pre-trained features refined to minimize the view-related information through adversarial training inspired by domain adaptation methods. Our method is able to recognize actions from unfamiliar viewpoints and works effectively on substantially less training data than the ones necessary to train state-of-the-art cross-view methods with exceptional results
Shearlets as Multi-scale Radon Transforms (STSIP)
No full textWe show that the 2D-shearlet transform is the composition of the a ne Radon transform, a 1D-wavelet transform and a 1D-convolution
HHP-Net: A light Heteroscedastic neural network for Head Pose estimation with uncertainty
No full textIn this paper we introduce a novel method to estimate the head pose of people in single images starting from a small set of head keypoints. To this purpose, we propose a regression model that exploits keypoints computed automatically by 2D pose estimation algorithms and outputs the head pose represented by yaw, pitch, and roll. Our model is simple to implement and more efficient with respect to the state of the art -faster in inference and smaller in terms of memory occupancy -with comparable accuracy.Our method also provides a measure of the heteroscedastic uncertainties associated with the three angles, through an appropriately designed loss function; we show there is a correlation between error and uncertainty values, thus this extra source of information may be used in subsequent computational steps. As an example application, we address social interaction analysis in images: we propose an algorithm for a quantitative estimation of the level of interaction between people, starting from their head poses and reasoning on their mutual positions
Cross-view action recognition with small-scale datasets
No full textCross-view action recognition refers to the task of recognizing actions observed from view-points that are unfamiliar to the system. To address the complexity of the problem, state of the art methods often rely on large-scale datasets, where the variability of viewpoints is appropriately represented. However, this comes to a significant price, in terms of computational power, time, costs, energy for both gathering data annotation and training the model. We propose a methodological pipeline that tackles the same challenges with specific focus on small-scale datasets and attention to the amount of resources required. The core idea of our method is to transfer knowledge from an intermediate, pre-trained representation, under the hypothesis that it already may implicitly incorporate relevant cues for the task. We rely on an effective domain adaptation strategy coupled with the design of a robust classifier that promotes view-invariant properties and allows us to efficiently generalise to action recognition to unseen viewpoints. In contrast to other state-of-art methods employing also alternative data modalities, our approach is purely video-based and thus has a wider field of applications. We present a thorough experimental analysis justifying the choices on the design of the pipeline, and providing a comparison with existing approaches in the two main scenarios of one-one learning and multiple view learning, where our approach provides superior performance
Knowledge distillation for efficient standard scanplane detection of fetal ultrasound
No full textAbstract: In clinical practice, ultrasound standard planes (SPs) selection is experience-dependent and it suffers from inter-observer and intra-observer variability. Automatic recognition of SPs can help improve the quality of examinations and make the evaluations more objective. In this paper, we propose a method for the automatic identification of SPs, to be installed onboard a portable ultrasound system with limited computational power. The deep Learning methodology we design is based on the concept of Knowledge Distillation, transferring knowledge from a large and well-performing teacher to a smaller student architecture. To this purpose, we evaluate a set of different potential teachers and students, as well as alternative knowledge distillation techniques, to balance a trade-off between performances and architectural complexity. We report a thorough analysis of fetal ultrasound data, focusing on a benchmark dataset, to the best of our knowledge the only one available to date. Graphical abstract: [Figure not available: see fulltext.]
Food Image Classification: The Benefit of In-Domain Transfer Learning
No full textMonitoring food intake and calories may be fundamental for a healthy lifestyle and preventing nutrition-related illnesses. Recently, deep-learning approaches have been extensively exploited to provide an automatic analysis of food images. However, food image datasets have peculiar challenges, including fine granularity with a high intra-class and low inter-class variability. In this work, we focus on training strategies considering the typical scenario where data availability and computational resources are limited. Exploiting convolutional neural networks, we show that in-domain source datasets provide a better representation with respect to only using ImageNet, bringing a significant increase in test accuracy. We finally show that ensembling different CNN models further improves the learned representation
- …
