Modeling visual attention-particularly stimulus-driven, saliency-based attention-has been a very active research area over the past 25 years. Many different models of attention are now available which, aside from lending theoretical contributions to other fields, have demonstrated successful applications in computer vision, mobile robotics, and cognitive systems. Here we review, from a computational perspective, the basic concepts of attention implemented in these models. We present a taxonomy of nearly 65 models, which provides a critical comparison of approaches, their capabilities, and shortcomings. In particular, 13 criteria derived from behavioral and computational studies are formulated for qualitative comparison of attention models. Furthermore, we address several challenging issues with models, including biological plausibility of the computations, correlation with eye movement datasets, bottom-up and top-down dissociation, and constructing meaningful performance measures. Finally, we highlight current research trends in attention modeling and provide insights for future.

State-of-the-Art in Visual Attention Modeling

We are on the verge of ubiquitously adopting Augmented Reality (AR) technologies to enhance our percep- tion and help us see, hear, and feel our environments in new and enriched ways. AR will support us in fields such as education, maintenance, design and reconnaissance, to name but a few. This paper describes the field of AR, including a brief definition and development history, the enabling technologies and their characteristics. It surveys the state of the art by reviewing some recent applications of AR technology as well as some known limitations regarding human factors in the use of AR systems that developers will need to overcome.

/pdf/a-survey-of-augmented-reality-technologies-applications-and-3d9hf1tnc4.pdf

A Survey of Augmented Reality Technologies, Applications and Limitations

Temporal relational reasoning, the ability to link meaningful transformations of objects or entities over time, is a fundamental property of intelligent species. In this paper, we introduce an effective and interpretable network module, the Temporal Relation Network (TRN), designed to learn and reason about temporal dependencies between video frames at multiple time scales. We evaluate TRN-equipped networks on activity recognition tasks using three recent video datasets - Something-Something, Jester, and Charades - which fundamentally depend on temporal relational reasoning. Our results demonstrate that the proposed TRN gives convolutional neural networks a remarkable capacity to discover temporal relations in videos. Through only sparsely sampled video frames, TRN-equipped networks can accurately predict human-object interactions in the Something-Something dataset and identify various human gestures on the Jester dataset with very competitive performance. TRN-equipped networks also outperform two-stream networks and 3D convolution networks in recognizing daily activities in the Charades dataset. Further analyses show that the models learn intuitive and interpretable visual common sense knowledge in videos (Code and models are available at http://relation.csail.mit.edu/.).

/pdf/temporal-relational-reasoning-in-videos-3v06umxf2r.pdf

Temporal Relational Reasoning in Videos

The explosive growth in video streaming gives rise to challenges on performing video understanding at high accuracy and low computation cost. Conventional 2D CNNs are computationally cheap but cannot capture temporal relationships; 3D CNN based methods can achieve good performance but are computationally intensive, making it expensive to deploy. In this paper, we propose a generic and effective Temporal Shift Module (TSM) that enjoys both high efficiency and high performance. Specifically, it can achieve the performance of 3D CNN but maintain 2D CNN’s complexity. TSM shifts part of the channels along the temporal dimension; thus facilitate information exchanged among neighboring frames. It can be inserted into 2D CNNs to achieve temporal modeling at zero computation and zero parameters. We also extended TSM to online setting, which enables real-time low-latency online video recognition and video object detection. TSM is accurate and efficient: it ranks the first place on the Something-Something leaderboard upon publication; on Jetson Nano and Galaxy Note8, it achieves a low latency of 13ms and 35ms for online video recognition. The code is available at: https://github. com/mit-han-lab/temporal-shift-module.

/pdf/tsm-temporal-shift-module-for-efficient-video-understanding-20eaekcf76.pdf

TSM: Temporal Shift Module for Efficient Video Understanding

This paper introduces a video dataset of spatio-temporally localized Atomic Visual Actions (AVA). The AVA dataset densely annotates 80 atomic visual actions in 437 15-minute video clips, where actions are localized in space and time, resulting in 1.59M action labels with multiple labels per person occurring frequently. The key characteristics of our dataset are: (1) the definition of atomic visual actions, rather than composite actions; (2) precise spatio-temporal annotations with possibly multiple annotations for each person; (3) exhaustive annotation of these atomic actions over 15-minute video clips; (4) people temporally linked across consecutive segments; and (5) using movies to gather a varied set of action representations. This departs from existing datasets for spatio-temporal action recognition, which typically provide sparse annotations for composite actions in short video clips. AVA, with its realistic scene and action complexity, exposes the intrinsic difficulty of action recognition. To benchmark this, we present a novel approach for action localization that builds upon the current state-of-the-art methods, and demonstrates better performance on JHMDB and UCF101-24 categories. While setting a new state of the art on existing datasets, the overall results on AVA are low at 15.8% mAP, underscoring the need for developing new approaches for video understanding.

/pdf/ava-a-video-dataset-of-spatio-temporally-localized-atomic-rqvc1zcfc8.pdf

AVA: A Video Dataset of Spatio-Temporally Localized Atomic Visual Actions

Neural networks trained on datasets such as ImageNet have led to major advances in visual object classification. One obstacle that prevents networks from reasoning more deeply about complex scenes and situations, and from integrating visual knowledge with natural language, like humans do, is their lack of common sense knowledge about the physical world. Videos, unlike still images, contain a wealth of detailed information about the physical world. However, most labelled video datasets represent high-level concepts rather than detailed physical aspects about actions and scenes. In this work, we describe our ongoing collection of the “something-something” database of video prediction tasks whose solutions require a common sense understanding of the depicted situation. The database currently contains more than 100,000 videos across 174 classes, which are defined as caption-templates. We also describe the challenges in crowd-sourcing this data at scale.

The “Something Something” Video Database for Learning and Evaluating Visual Common Sense

Gesture recognition and its application in human-computer interfaces have been growing increasingly popular in recent years. Although many gestures can be recognized from a single image frame, to build a responsive, accurate system, that can recognize complex gestures with subtle differences between them we need large-scale real-world video datasets. In this work, we introduce the largest collection of short clips of videos of humans performing gestures in front of the camera. The dataset has been collected with the help of over 1300 different actors in their unconstrained environments. Additionally, we present an on-going gesture recognition challenge based on our dataset and the current results. We also describe how a baseline achieving over 93% recognition accuracy can be obtained with a simple 3D convolutional neural network.

/pdf/the-jester-dataset-a-large-scale-video-dataset-of-human-3fb1japm9l.pdf

The Jester Dataset: A Large-Scale Video Dataset of Human Gestures

We describe an augmented reality system designed for online acquisition of visual knowledge and retrieval of memorized objects. The system relies on a head mounted camera and display, which allow the user to view the environment together with overlaid augmentations by the system. In this setup, communication by hand gestures and speech is mandatory as common input devices like mouse and keyboard are not available. Using gesture and speech, basically three types of tasks must be handled: (i) Communication with the system about the environment, in particular, directing attention towards objects and commanding the memorization of sample views; (ii) control of system operation, e.g. switching between display modes; and (iii) re-adaptation of the interface itself in case communication becomes unreliable due to changes in external factors, such as illumination conditions. We present an architecture to manage these tasks and describe and evaluate several of its key elements, including modules for pointing gesture recognition, menu control based on gesture and speech, and control strategies to cope with situations when vision becomes unreliable and has to be re-adapted by speech.

/pdf/multimodal-interaction-in-an-augmented-reality-scenario-4htxv477ka.pdf

Multimodal interaction in an augmented reality scenario

The “life” of most neural vision systems splits into a one-time training phase and an application phase during which knowledge is no longer acquired. This is both technically inflexible and cognitively unsatisfying. Here we propose an appearance based vision system for object recognition which can be adapted online, both to acquire visual knowledge about new objects and to correct erroneous classification. The system works in an office scenario, acquisition of object knowledge is triggered by hand gestures. The neural classifier offers two ways of training: Firstly, the new samples can be added immediately to the classifier to obtain a running system at once, though at the cost of reduced classification performance. Secondly, a parallel processing branch adapts the classification system thoroughly to the enlarged image domain and loads the new classifier to the running system when ready.

/pdf/adaptive-computer-vision-online-learning-for-object-26su2t3rfa.pdf

Adaptive Computer Vision: Online Learning for Object Recognition

We present a vision system for human-machine interaction that relies on a small wearable camera which can be mounted to common glasses The camera views the area in front of the user, especially the hands To evaluate hand movements for pointing gestures to objects and to recognise object reference, an approach relying on the integration of bottom-up generated feature maps and top-down propagated recognition results is introduced In this vision system, modules for context free focus of attention work in parallel to a recognition system for hand gestures In contrast to other approaches, the fusion of the two branches is not on the symbolic but on the sub-symbolic level by use of attention maps This method is plausible from a cognitive point of view and facilitates the integration of entirely different modalities

Ingo Bax

Papers

The “Something Something” Video Database for Learning and Evaluating Visual Common Sense

The Jester Dataset: A Large-Scale Video Dataset of Human Gestures

Multimodal interaction in an augmented reality scenario

Adaptive Computer Vision: Online Learning for Object Recognition

Integrating context-free and context-dependent attentional mechanisms for gestural object reference