3M-TRANSFORMER: A Multi-Stage Multi-Stream Multimodal Transformer for Embodied Turn-Taking Prediction

Predicting turn-taking in multiparty conversations has many practical applications in human-computer/robot interaction. However, the complexity of human communication makes it a challenging task. Recent advances have shown that synchronous multi-perspective egocentric data can significantly improve turn-taking prediction compared to asynchronous, single-perspective transcriptions. Building on this research, we propose a new multimodal transformer-based architecture for predicting turn-taking in embodied, synchronized multi-perspective data. Our experimental results on the recently introduced EgoCom dataset show a substantial performance improvement of up to 14.01% on average compared to existing baselines and alternative transformer-based approaches. The source code, and the pre-trained models of our 3M-Transformer will be available upon acceptance.

Debiased-CAM to mitigate systematic error with faithful visual explanations of machine learning

Model explanations such as saliency maps can improve user trust in AI by highlighting important features for a prediction. However, these become distorted and misleading when explaining predictions of images that are subject to systematic error (bias). Furthermore, the distortions persist despite model fine-tuning on images biased by different factors (blur, color temperature, day/night). We present Debiased-CAM to recover explanation faithfulness across various bias types and levels by training a multi-input, multi-task model with auxiliary tasks for explanation and bias level predictions. In simulation studies, the approach not only enhanced prediction accuracy, but also generated highly faithful explanations about these predictions as if the images were unbiased. In user studies, debiased explanations improved user task performance, perceived truthfulness and perceived helpfulness. Debiased training can provide a versatile platform for robust performance and explanation faithfulness for a wide range of applications with data biases.

Enhancing Egocentric 3D Pose Estimation with Third Person Views

In this paper, we propose a novel approach to enhance the 3D body pose estimation of a person computed from videos captured from a single wearable camera. The key idea is to leverage high-level features linking first- and third-views in a joint embedding space. To learn such embedding space we introduce First2Third-Pose, a new paired synchronized dataset of nearly 2,000 videos depicting human activities captured from both first- and third-view perspectives. We explicitly consider spatial- and motion-domain features, combined using a semi-Siamese architecture trained in a self-supervised fashion. Experimental results demonstrate that the joint multi-view embedded space learned with our dataset is useful to extract discriminatory features from arbitrary single-view egocentric videos, without needing domain adaptation nor knowledge of camera parameters. We achieve significant improvement of egocentric 3D body pose estimation performance on two unconstrained datasets, over three supervised state-of-the-art approaches. Our dataset and code will be available for research purposes.

Learning grounded word meaning representations on similarity graphs

This paper introduces a novel approach to learn visually grounded meaning representations of words as low-dimensional node embeddings on an underlying graph hierarchy. The lower level of the hierarchy models modality-specific word representations through dedicated but communicating graphs, while the higher level puts these representations together on a single graph to learn a representation jointly from both modalities. The topology of each graph models similarity relations among words, and is estimated jointly with the graph embedding. The assumption underlying this model is that words sharing similar meaning correspond to communities in an underlying similarity graph in a low-dimensional space. We named this model Hierarchical Multi-Modal Similarity Graph Embedding (HM-SGE). Experimental results validate the ability of HM-SGE to simulate human similarity judgements and concept categorization, outperforming the state of the art.

Graph Constrained Data Representation Learning for Human Motion Segmentation

Recently, transfer subspace learning based approaches have shown to be a valid alternative to unsupervised subspace clustering and temporal data clustering for human motion segmentation (HMS). These approaches leverage prior knowledge from a source domain to improve clustering performance on a target domain, and currently they represent the state of the art in HMS. Bucking this trend, in this paper, we propose a novel unsupervised model that learns a representation of the data and digs clustering information from the data itself. Our model is reminiscent of temporal subspace clustering, but presents two critical differences. First, we learn an auxiliary data matrix that can deviate from the initial data, hence confer more degrees of freedom to the coding matrix. Second, we introduce a regularization term for this auxiliary data matrix that preserves the local geometrical structure present in the high-dimensional space. The proposed model is efficiently optimized by using an original Alternating Direction Method of Multipliers (ADMM) formulation allowing to learn jointly the auxiliary data representation, a nonnegative dictionary and a coding matrix. Experimental results on four benchmark datasets for HMS demonstrate that our approach achieves significantly better clustering performance then state-of-the-art methods, including both unsupervised and more recent semi-supervised transfer learning approaches.

Interaction-GCN: a Graph Convolutional Network based framework for social interaction recognition in egocentric videos

In this paper we propose a new framework to categorize social interactions in egocentric videos, we named InteractionGCN. Our method extracts patterns of relational and non-relational cues at the frame level and uses them to build a relational graph from which the interactional context at the frame level is estimated via a Graph Convolutional Network based approach. Then it propagates this context over time, together with first-person motion information, through a Gated Recurrent Unit architecture. Ablation studies and experimental evaluation on two publicly available datasets validate the proposed approach and establish state of the art results.

Modeling long-term interactions to enhance action recognition

In this paper, we propose a new approach to under-stand actions in egocentric videos that exploits the semantics of object interactions at both frame and temporal levels. At the frame level, we use a region-based approach that takes as input a primary region roughly corresponding to the user hands and a set of secondary regions potentially corresponding to the interacting objects and calculates the action score through a CNN formulation. This information is then fed to a Hierarchical LongShort-Term Memory Network (HLSTM) that captures temporal dependencies between actions within and across shots. Ablation studies thoroughly validate the proposed approach, showing in particular that both levels of the HLSTM architecture contribute to performance improvement. Furthermore, quantitative comparisons show that the proposed approach outperforms the state-of-the-art in terms of action recognition on standard benchmarks,without relying on motion information

Debiased-CAM for bias-agnostic faithful visual explanations of deep convolutional networks

Class activation maps (CAMs) explain convolutional neural network predictions by identifying salient pixels, but they become misaligned and misleading when explaining predictions on images under bias, such as images blurred accidentally or deliberately for privacy protection, or images with improper white balance. Despite model fine-tuning to improve prediction performance on these biased images, we demonstrate that CAM explanations become more deviated and unfaithful with increased image bias. We present Debiased-CAM to recover explanation faithfulness across various bias types and levels by training a multi-input, multi-task model with auxiliary tasks for CAM and bias level predictions. With CAM as a prediction task, explanations are made tunable by retraining the main model layers and made faithful by self-supervised learning from CAMs of unbiased images. The model provides representative, bias-agnostic CAM explanations about the predictions on biased images as if generated from their unbiased form. In four simulation studies with different biases and prediction tasks, Debiased-CAM improved both CAM faithfulness and task performance. We further conducted two controlled user studies to validate its truthfulness and helpfulness, respectively. Quantitative and qualitative analyses of participant responses confirmed Debiased-CAM as more truthful and helpful. Debiased-CAM thus provides a basis to generate more faithful and relevant explanations for a wide range of real-world applications with various sources of bias.

Seeing and Hearing Egocentric Actions: How Much Can We Learn?

Our interaction with the world is an inherently multimodal experience. However, the understanding of human-to-object interactions has historically been addressed focusing on a single modality. In particular, a limited number of works have considered to integrate the visual and audio modalities for this purpose. In this work, we propose a multimodal approach for egocentric action recognition in a kitchen environment that relies on audio and visual information. Our model combines a sparse temporal sampling strategy with a late fusion of audio, spatial, and temporal streams. Experimental results on the EPIC-Kitchens dataset show that multimodal integration leads to better performance than unimodal approaches. In particular, we achieved a 5.18% improvement over the state of the art on verb classification.

Learning event representations in image sequences by dynamic graph embedding

Recently, self-supervised learning has proved to be effective to learn representations of events in image sequences, where events are understood as sets of temporally adjacent images that are semantically perceived as a whole. However, although this approach does not require expensive manual annotations, it is data hungry and suffers from domain adaptation problems. As an alternative, in this work, we propose a novel approach for learning event representations named Dynamic Graph Embedding (DGE). The assumption underlying our model is that a sequence of images can be represented by a graph that encodes both semantic and temporal similarity. The key novelty of DGE is to learn jointly the graph and its graph embedding. At its core, DGE works by iterating over two steps: 1) updating the graph representing the semantic and temporal structure of the data based on the current data representation, and 2) updating the data representation to take into account the current data graph structure. The main advantage of DGE over state-of-the-art self-supervised approaches is that it does not require any training set, but instead learns iteratively from the data itself a low-dimensional embedding that reflects their temporal and semantic structure. Experimental results on two benchmark datasets of real image sequences captured at regular intervals demonstrate that the proposed DGE leads to effective event representations. In particular, it achieves robust temporal segmentation on the EDUBSeg and EDUBSeg-Desc benchmark datasets, outperforming the state of the art.