This paper accelerates video perception, such as segmentation and human pose estimation, by levering cross-frame redundancies. Unlike the existing approaches, which avoid redundant computations by warping the past features using optical-flow or by performing sparse convolutions on frame differences, we approach the problem from a different perspective: low-bit quantization. We observe that residuals, as the difference in network activations between two neighboring frames, exhibit properties that make them highly quantizable. Based on this observation, we propose a novel quantization scheme for video networks coined as Res idual Quantization. ResQ extends the standard, frame-by-frame, quantization scheme by incorporating temporal dependencies that lead to better performance in terms of accuracy vs. bit-width. Furthermore, we extend our model to dynamically adjust the bit-width proportionally to the amount of changes in the video. We showcase the superiority of our model, against the standard quantization and existing efficient video perception models, using various architectures on semantic segmentation, video object segmentation and human pose estimation benchmarks.

We propose Skip-Convolutions to leverage the large amount of redundancies in video streams and save computations. Each video is represented as a series of changes across frames and network activations, denoted as residuals. We reformulate standard convolution to be efficiently computed on residual frames: each layer is coupled with a binary gate deciding whether a residual is important to the model prediction, e.g. foreground regions, or it can be safely skipped, e.g. background regions. These gates can either be implemented as an efficient network trained jointly with convolution kernels, or can simply skip the residuals based on their magnitude. Gating functions can also incorporate block-wise sparsity structures, as required for efficient implementation on hardware platforms. By replacing all convolutions with Skip-Convolutions in two state-of-the-art architectures, namely EfficientDet and HRNet, we reduce their computational cost consistently by a factor of 3 ∼ 4× for two different tasks, without any accuracy drop. Extensive comparisons with existing model compression, as well as image and video efficiency methods demonstrate that Skip-Convolutions set a new state-of-the-art by effectively exploiting the temporal redundancies in videos.

Convolutional Neural Networks experience catastrophic forgetting when optimized on a sequence of learning problems: as they meet the objective of the current training examples, their performance on previous tasks drops drastically. In this work, we introduce a novel framework to tackle this problem with conditional computation. We equip each convolutional layer with task-specific gating modules, selecting which filters to apply on the given input. This way, we achieve two appealing properties. Firstly, the execution patterns of the gates allow to identify and protect important filters, ensuring no loss in the performance of the model for previously learned tasks. Secondly, by using a sparsity objective, we can promote the selection of a limited set of kernels, allowing to retain sufficient model capacity to digest new tasks. Existing solutions require, at test time, awareness of the task to which each example belongs to. This knowledge, however, may not be available in many practical scenarios. Therefore, we additionally introduce a task classifier that predicts the task label of each example, to deal with settings in which a task oracle is not available. We validate our proposal on four continual learning datasets. Results show that our model consistently outperforms existing methods both in the presence and the absence of a task oracle. Notably, on Split SVHN and Imagenet-50 datasets, our model yields up to 23.98% and 17.42% improvement in accuracy w.r.t. competing methods.

Continual Learning has inspired a plethora of approaches and evaluation settings; however, the majority of them overlooks the properties of a practical scenario, where the data stream cannot be shaped as a sequence of tasks and offline training is not viable. We work towards General Continual Learning (GCL), where task boundaries blur and the domain and class distributions shift either gradually or suddenly. We address it through mixing rehearsal with knowledge distillation and regularization; our simple baseline, Dark Experience Replay, matches the network's logits sampled throughout the optimization trajectory, thus promoting consistency with its past. By conducting an extensive analysis on both standard benchmarks and a novel GCL evaluation setting (MNIST-360), we show that such a seemingly simple baseline outperforms consolidated approaches and leverages limited resources. We further explore the generalization capabilities of our objective, showing its regularization being beneficial beyond mere performance.

We present a novel and hierarchical approach for supervised classification of signals spanning over a fixed graph, reflecting shared properties of the dataset. To this end, we introduce a Convolutional Cluster Pooling layer exploiting a multi-scale clustering in order to highlight, at different resolutions, locally connected regions on the input graph. Our proposal generalises well-established neural models such as Convolutional Neural Networks (CNNs) on irregular and complex domains, by means of the exploitation of the weight sharing property in a graph-oriented architecture. In this work, such property is based on the centrality of each vertex within its soft-assigned cluster. Extensive experiments on NTU RGB+D, CIFAR-10 and 20NEWS demonstrate the effectiveness of the proposed technique in capturing both local and global patterns in graph-structured data out of different domains.

Novelty detection is commonly referred as the discrimination of observations that do not conform to a learned model of regularity. Despite its importance in different application settings, designing a novelty detector is utterly complex due to the unpredictable nature of novelties and its inaccessibility during the training procedure, factors which expose the unsupervised nature of the problem. In our proposal, we design a general unsupervised framework where we equip a deep autoencoder with a parametric density estimator that learns the probability distribution underlying the latent representations with an autoregressive procedure. We show that a maximum likelihood objective, optimized in conjunction with the reconstruction of normal samples, effectively acts as a regularizer for the task at hand, by minimizing the differential entropy of the distribution spanned by latent vectors. In addition to providing a very general formulation, extensive experiments of our model on publicly available datasets deliver on-par or superior performances if compared to state-of-the-art methods in one-class and in video anomaly detection settings. Differently from our competitors, we remark that our proposal does not make any assumption about the nature of the novelties, making our work easily applicable to disparate contexts.

In this work we aim to predict the driver's focus of attention. The goal is to estimate what a person would pay attention to while driving, and which part of the scene around the vehicle is more critical for the task. To this end we propose a new computer vision model based on a multi-branch deep architecture that integrates three sources of information: raw video, motion and scene semantics. We also introduce DR(eye)VE, the largest dataset of driving scenes for which eye-tracking annotations are available. This dataset features more than 500,000 registered frames, matching ego-centric views (from glasses worn by drivers) and car-centric views (from roof-mounted camera), further enriched by other sensors measurements. Results highlight that several attention patterns are shared across drivers and can be reproduced to some extent. The indication of which elements in the scene are likely to capture the driver's attention may benefit several applications in the context of human-vehicle interaction and driver attention analysis.

Dense optical flow estimation is complex and time consuming, with state-of-the-art methods relying either on large synthetic data sets or on pipelines requiring up to a few minutes per frame pair. In this paper, we address the problem of optical flow estimation in the automotive scenario in a self-supervised manner. We argue that optical flow can be cast as a geometrical warping between two successive video frames and devise a deep architecture to estimate such transformation in two stages. First, a dense pixel-level flow is computed with a projective bootstrap on rigid surfaces. We show how such global transformation can be approximated with a homography and extend spatial transformer layers so that they can be employed to compute the flow field implied by such transformation. Subsequently, we refine the prediction by feeding a second, deeper network that accounts for moving objects. A final reconstruction loss compares the warping of frame Xtwith the subsequent frame Xt+1and guides both estimates. The model has the speed advantages of end-to-end deep architectures while achieving competitive performances, both outperforming recent unsupervised methods and showing good generalization capabilities on new automotive data sets.