Carl Vondrick

A large-scale benchmark dataset for event recognition in surveillance video

We introduce a new large-scale video dataset designed to assess the performance of diverse visual event recognition algorithms with a focus on continuous visual event recognition (CVER) in outdoor areas with wide coverage. Previous datasets for action recognition are unrealistic for real-world surveillance because they consist of short clips showing one action by one individual [15, 8]. Datasets have been developed for movies [11] and sports [12], but, these actions and scene conditions do not apply effectively to surveillance videos. Our dataset consists of many outdoor scenes with actions occurring naturally by non-actors in continuously captured videos of the real world. The dataset includes large numbers of instances for 23 event types distributed throughout 29 hours of video. This data is accompanied by detailed annotations which include both moving object tracks and event examples, which will provide solid basis for large-scale evaluation. Additionally, we propose different types of evaluation modes for visual recognition tasks and evaluation metrics along with our preliminary experimental results. We believe that this dataset will stimulate diverse aspects of computer vision research and help us to advance the CVER tasks in the years ahead.

Efficiently Scaling up Crowdsourced Video Annotation

We present an extensive three year study on economically annotating video with crowdsourced marketplaces. Our public framework has annotated thousands of real world videos, including massive data sets unprecedented for their size, complexity, and cost. To accomplish this, we designed a state-of-the-art video annotation user interface and demonstrate that, despite common intuition, many contemporary interfaces are sub-optimal. We present several user studies that evaluate different aspects of our system and demonstrate that minimizing the cognitive load of the user is crucial when designing an annotation platform. We then deploy this interface on Amazon Mechanical Turk and discover expert and talented workers who are capable of annotating difficult videos with dense and closely cropped labels. We argue that video annotation requires specialized skill; most workers are poor annotators, mandating robust quality control protocols. We show that traditional crowdsourced micro-tasks are not suitable for video annotation and instead demonstrate that deploying time-consuming macro-tasks on MTurk is effective. Finally, we show that by extracting pixel-based features from manually labeled key frames, we are able to leverage more sophisticated interpolation strategies to maximize performance given a fixed budget. We validate the power of our framework on difficult, real-world data sets and we demonstrate an inherent trade-off between the mix of human and cloud computing used vs. the accuracy and cost of the labeling. We further introduce a novel, cost-based evaluation criteria that compares vision algorithms by the budget required to achieve an acceptable performance. We hope our findings will spur innovation in the creation of massive labeled video data sets and enable novel data-driven computer vision applications.

Do We Need More Training Data or Better Models for Object Detection

Datasets for training object recognition systems are steadily growing in size. This paper investigates the question of whether existing detectors will continue to improve as data grows, or if models are close to saturating due to limited model complexity and the Bayes risk associated with the feature spaces in which they operate. We focus on the popular paradigm of scanning-window templates defined on oriented gradient features, trained with discriminative classifiers. We investigate the performance of mixtures of templates as a function of the number of templates (complexity) and the amount of training data. We find that additional data does help, but only with correct regularization and treatment of noisy examples or “outliers” in the training data. Surprisingly, the performance of problem domain-agnostic mixture models appears to saturate quickly (∼10 templates and ∼100 positive training examples per template). However, compositional mixtures (implemented via composed parts) give much better performance because they share parameters among templates, and can synthesize new templates not encountered during training. This suggests there is still room to improve performance with linear classifiers and the existing feature space by improved representations and learning algorithms.

Efficiently scaling up video annotation with crowdsourced marketplaces

Accurately annotating entities in video is labor intensive and expensive. As the quantity of online video grows, traditional solutions to this task are unable to scale to meet the needs of researchers with limited budgets. Current practice provides a temporary solution by paying dedicated workers to label a fraction of the total frames and otherwise settling for linear interpolation. As budgets and scale require sparser key frames, the assumption of linearity fails and labels become inaccurate. To address this problem we have created a public framework for dividing the work of labeling video data into micro-tasks that can be completed by huge labor pools available through crowdsourced marketplaces. By extracting pixel-based features from manually labeled entities, we are able to leverage more sophisticated interpolation between key frames to maximize performance given a budget. Finally, by validating the power of our framework on difficult, real-world data sets we demonstrate an inherent trade-off between the mix of human and cloud computing used vs. the accuracy and cost of the labeling.

Anticipating Visual Representations from Unlabeled Video

Anticipating actions and objects before they start or appear is a difficult problem in computer vision with several real-world applications. This task is challenging partly because it requires leveraging extensive knowledge of the world that is difficult to write down. We believe that a promising resource for efficiently learning this knowledge is through readily available unlabeled video. We present a framework that capitalizes on temporal structure in unlabeled video to learn to anticipate human actions and objects. The key idea behind our approach is that we can train deep networks to predict the visual representation of images in the future. Visual representations are a promising prediction target because they encode images at a higher semantic level than pixels yet are automatic to compute. We then apply recognition algorithms on our predicted representation to anticipate objects and actions. We experimentally validate this idea on two datasets, anticipating actions one second in the future and objects five seconds in the future.

Assessing the Quality of Actions

While recent advances in computer vision have provided reliable methods to recognize actions in both images and videos, the problem of assessing how well people perform actions has been largely unexplored in computer vision. Since methods for assessing action quality have many real-world applications in healthcare, sports, and video retrieval, we believe the computer vision community should begin to tackle this challenging problem. To spur progress, we introduce a learning-based framework that takes steps towards assessing how well people perform actions in videos. Our approach works by training a regression model from spatiotemporal pose features to scores obtained from expert judges. Moreover, our approach can provide interpretable feedback on how people can improve their action. We evaluate our method on a new Olympic sports dataset, and our experiments suggest our framework is able to rank the athletes more accurately than a non-expert human. While promising, our method is still a long way to rivaling the performance of expert judges, indicating that there is significant opportunity in computer vision research to improve on this difficult yet important task.

Learning Aligned Cross-Modal Representations from Weakly Aligned Data

People can recognize scenes across many different modalities beyond natural images. In this paper, we investigate how to learn cross-modal scene representations that transfer across modalities. To study this problem, we introduce a new cross-modal scene dataset. While convolutional neural networks can categorize cross-modal scenes well, they also learn an intermediate representation not aligned across modalities, which is undesirable for crossmodal transfer applications. We present methods to regularize cross-modal convolutional neural networks so that they have a shared representation that is agnostic of the modality. Our experiments suggest that our scene representation can help transfer representations across modalities for retrieval. Moreover, our visualizations suggest that units emerge in the shared representation that tend to activate on consistent concepts independently of the modality.

Generating the Future with Adversarial Transformers

We learn models to generate the immediate future in video. This problem has two main challenges. Firstly, since the future is uncertain, models should be multi-modal, which can be difficult to learn. Secondly, since the future is similar to the past, models store low-level details, which complicates learning of high-level semantics. We propose a framework to tackle both of these challenges. We present a model that generates the future by transforming pixels in the past. Our approach explicitly disentangles the models memory from the prediction, which helps the model learn desirable invariances. Experiments suggest that this model can generate short videos of plausible futures. We believe predictive models have many applications in robotics, health-care, and video understanding.

Cross-Modal Scene Networks

People can recognize scenes across many different modalities beyond natural images. In this paper, we investigate how to learn cross-modal scene representations that transfer across modalities. To study this problem, we introduce a new cross-modal scene dataset. While convolutional neural networks can categorize scenes well, they also learn an intermediate representation not aligned across modalities, which is undesirable for cross-modal transfer applications. We present methods to regularize cross-modal convolutional neural networks so that they have a shared representation that is agnostic of the modality. Our experiments suggest that our scene representation can help transfer representations across modalities for retrieval. Moreover, our visualizations suggest that units emerge in the shared representation that tend to activate on consistent concepts independently of the modality.

Do We Need More Training Data

Datasets for training object recognition systems are steadily increasing in size. This paper investigates the question of whether existing detectors will continue to improve as data grows, or saturate in performance due to limited model complexity and the Bayes risk associated with the feature spaces in which they operate. We focus on the popular paradigm of discriminatively trained templates defined on oriented gradient features. We investigate the performance of mixtures of templates as the number of mixture components and the amount of training data grows. Surprisingly, even with proper treatment of regularization and “outliers”, the performance of classic mixture models appears to saturate quickly (