Open Access

Efficient algorithms for object recognition are crucial for the newly robotics and computer vision applications that demand real-time and on-line methods. Some examples are autonomous systems, navigating robots, autonomous driving. In this work, we focus on efficient semantic segmentation, which is the problem of labeling each pixel of an image with a semantic class. Our aim is to speed-up all of the parts of the semantic segmentation pipeline. We also aim at delivering a labeling solution on a time budget, that can be decided on-the-fly. For this purpose, we analyze all the components of the semantic segmentation pipeline, and identify the computational bottleneck of each of them. The different components of the pipeline are over-segmenting the image with local regions, extracting features and classify the local regions, and the final inference of the image labeling with semantic classes. We focus on each of these steps. First, we introduce a new superpixel algorithm to over-segment the image. Our superpixel method runs in real-time and can deliver a solution at any time budget. Then, for feature extraction, we focus on the framework that computes descriptors and encodes them, followed by a pooling step. We see that the encoding step is the bottleneck, for computational efficiency and performance. We present a novel assignment-based encoding formulation, that allows for the design of a new, very efficient, encoding. Finally, the image labeling output is obtained modeling the dependencies with a Conditional Random Field (CRF). In semantic image segmentation, the computational cost of instantiating the potentials is much higher than MAP inference. We introduce Active MAP inference to on-the-fly select a subset of potentials to be instantiated in the energy function, leaving the rest as unknown, and to estimate the MAP labeling from such incomplete energy function. We perform experiments on all proposed methods for the different parts of the semantic segmentation pipeline. We show that our superpixel extraction achieves higher accuracy than state-of-the-art on standard superpixel benchmark, while it runs in real-time. We test our feature encoding on standard image classification and segmentation benchmarks, and we show that our method achieves competitive results with the state-of-the-art, and requires less time and memory. Finally, results for semantic segmentation benchmark show that Active MAP inference achieves similar levels of accuracy but with major efficiency gains.

Grasping the meaning of everyday visual events is a fundamental feat of human intelligence that hinges on diverse neural processes ranging from vision to higher-level cognition. Deciphering the neural basis of visual event understanding requires rich, extensive, and appropriately designed experimental data. However, this type of data is hitherto missing. To fill this gap, we introduce the BOLD Moments Dataset (BMD), a large dataset of whole-brain fMRI responses to over 1,000 short (3s) naturalistic video clips and accompanying metadata. We show visual events interface with an array of processes, extending even to memory, and we reveal a match in hierarchical processing between brains and video-computable deep neural networks. Furthermore, we showcase that BMD successfully captures temporal dynamics of visual events at second resolution. BMD thus establishes a critical groundwork for investigations of the neural basis of visual event understanding.

Visual scene perception is mediated by a set of cortical regions that respond preferentially to images of scenes, including the occipital place area (OPA) and parahippocampal place area (PPA). However, the differential contribution of OPA and PPA to scene perception remains an open research question. In this study, we take a deep neural network (DNN)-based computational approach to investigate the differences in OPA and PPA function. In a first step we search for a computational model that predicts fMRI responses to scenes in OPA and PPA well. We find that DNNs trained to predict scene components (e.g., wall, ceiling, floor) explain higher variance uniquely in OPA and PPA than a DNN trained to predict scene category (e.g., bathroom, kitchen, office). This result is robust across several DNN architectures. On this basis, we then determine whether particular scene components predicted by DNNs differentially account for unique variance in OPA and PPA. We find that variance in OPA responses uniquely explained by the navigation-related floor component is higher compared to the variance explained by the wall and ceiling components. In contrast, PPA responses are better explained by the combination of wall and floor, that is scene components that together contain the structure and texture of the scene. This differential sensitivity to scene components suggests differential functions of OPA and PPA in scene processing. Moreover, our results further highlight the potential of the proposed computational approach as a general tool in the investigation of the neural basis of human scene perception.

In this paper, we introduce an approach for future frames prediction based on a single input image. Our method is able to generate an entire video sequence based on the information contained in the input frame. We adopt an autoregressive approach in our generation process, i.e., the output from each time step is fed as the input to the next step. Unlike other video prediction methods that use “one shot” generation, our method is able to preserve much more details from the input image, while also capturing the critical pixel-level changes between the frames. We overcome the problem of generation quality degradation by introducing a “complementary mask” module in our architecture, and we show that this allows the model to only focus on the generation of the pixels that need to be changed, and to reuse those that should remain static from its previous frame. We empirically validate our methods against various video prediction models on the UT Dallas Dataset, and show that our approach is able to generate high quality realistic video sequences from one static input image. In addition, we also validate the robustness of our method by testing a pre-trained model on the unseen ADFES facial expression dataset. We also provide qualitative results of our model tested on a human action dataset: The Weizmann Action database.

In this paper, we tackle the problem of predicting the affective responses of movie viewers, based on the content of the movies. Current studies on this topic focus on video representation learning and fusion techniques to combine the extracted features for predicting affect. Yet, these typically, while ignoring the correlation between multiple modality inputs, ignore the correlation between temporal inputs (i.e., sequential features). To explore these correlations, a neural network architecture—namely AttendAffectNet (AAN)—uses the self-attention mechanism for predicting the emotions of movie viewers from different input modalities. Particularly, visual, audio, and text features are considered for predicting emotions (and expressed in terms of valence and arousal). We analyze three variants of our proposed AAN: Feature AAN, Temporal AAN, and Mixed AAN. The Feature AAN applies the self-attention mechanism in an innovative way on the features extracted from the different modalities (including video, audio, and movie subtitles) of a whole movie to, thereby, capture the relationships between them. The Temporal AAN takes the time domain of the movies and the sequential dependency of affective responses into account. In the Temporal AAN, self-attention is applied on the concatenated (multimodal) feature vectors representing different subsequent movie segments. In the Mixed AAN, we combine the strong points of the Feature AAN and the Temporal AAN, by applying self-attention first on vectors of features obtained from different modalities in each movie segment and then on the feature representations of all subsequent (temporal) movie segments. We extensively trained and validated our proposed AAN on both the MediaEval 2016 dataset for the Emotional Impact of Movies Task and the extended COGNIMUSE dataset. Our experiments demonstrate that audio features play a more influential role than those extracted from video and movie subtitles when predicting the emotions of movie viewers on these datasets. The models that use all visual, audio, and text features simultaneously as their inputs performed better than those using features extracted from each modality separately. In addition, the Feature AAN outperformed other AAN variants on the above-mentioned datasets, highlighting the importance of taking different features as context to one another when fusing them. The Feature AAN also performed better than the baseline models when predicting the valence dimension.