Efficient Depth-based Deep Learning Methods for Multi-Party Pose Estimation

Human detection and pose estimation are essential components for any artificial system responsive to the presence of humans and that react according to human-centered tasks. Robotic systems are typical examples, for which the body pose represents fine grained information useful to understand the behavior and activities of people, and interact with them. However, it is a challenging research topic with increasing difficulty given the unknown number of people in a usual scenario and factors like occlusions and sensing conditions. Current state-of-the-art methods have largely used deep Convolutional Neural Networks (CNN) to address the task. Traditionally, the selected CNNs are very deep and overparameterized, hence requiring large amounts of data to achieve good generalization and prevent overfitting. As a consequence, they are not straightforward to deploy in the low budget hardware typically available in practical applications such as HRI.This thesis studies methods for efficient and reliable 2D and 3D human pose estimation using deep learning approaches. It investigates novel lightweight convolutional network architectures that achieve real-time performance in multi-person scenarios and explores knowledge distillation methods to boost the performance of these models while keeping their efficiency. Moreover, this thesis addresses the high cost of data collection with annotations that arises with our deep learning-based approaches by relying on a large scale dataset of synthetic images with high variability. Domain adaptation methods and data augmentation strategies are proposed to exploit the synthetic corpus in order to achieve good generalization in sensor data. Additionally, this dissertation studies human 3D motion prediction framed as a sequence-to-sequence problem. Non-autoregressive transformer neural networks are proposed to predict elements in parallel to avoid error propagation from predicted elements, observed in autoregressive methods, while at the same time being efficient.Overall this thesis proposes different efficient and accurate deep learning solutions to design components of a human behaviour understanding system exploited in Human-Robot-Interaction (HRI) scenarios.

Word Sequence Modeling using Deep Learning

Joël Yvon Roland Legrand

For a long time, natural language processing (NLP) has relied on generative models with task specific and manually engineered features. Recently, there has been a resurgence of interest for neural networks in the machine learning community, obtaining state-of-the-art results in various fields such as computer vision, speech processing and natural language processing. The central idea behind these approaches is to learn features and models simultaneously, in an end-to-end manner, and making as few assumptions as possible. In NLP, word embeddings, mapping words in a dictionary on a continuous low-dimensional vector space, have proven to be very efficient for a large variety of tasks while requiring almost no a-priori linguistic assumptions. In this thesis, we investigate continuous representations of segments in a sentence for the purpose of solving NLP tasks that involve complex sentence-level relationships. Our sequence modelling approach is based on neural networks and takes advantage of word embeddings. A first approach models words in context in the form of continuous vector representations which are used to solve the task of interest. With the use of a compositional procedure, allowing arbitrarily-sized segments to be compressed onto continuous vectors, the model is able to consider long-range word dependencies as well. We first validate our approach on the task of bilingual word alignment, consisting in finding word correspondences between a sentence in two different languages. Source and target words in context are modeled using convolutional neural networks, obtaining representations that are later used to compute alignment scores. An aggregation operation enables unsupervised training for this task. We show that our model outperforms a standard generative model. The model above is extended to tackle phrase prediction tasks where phrases rather than single words are to be tagged. These tasks have been typically cast as classic word tagging problems using special tagging schemes to identify the segments boundaries. The proposed neural model focuses on learning fixed-size representations of arbitrarily-sized chunks of words that are used to solve the tagging task. A compositional operation is introduced in this work for the purpose of computing these representations. We demonstrate the viability of the proposed representations by evaluating the approach on the multiwork expression tagging task. The remainder of this thesis addresses the task of syntactic constituency parsing which, as opposed to the above tasks, aims at producing a structured output, in the form of a tree, of an input sentence. Syntactic parsing is cast as multiple phrase prediction problems that are solved recursively in a greedy manner. An extension using recursive compositional vector representations, allowing for lexical infor- mation to be propagated from early stages, is explored as well. This approach is evaluated on a standard corpus obtaining performance comparable to generative models with much shorter computation time. Finally, morphological tags are included as additional features, using a similar composition procedure, to improve the parsing performance for morphologically rich languages. State-of-the-art results were obtained for these task and languages.

EPFL2016

Large-Scale Image Segmentation with Convolutional Networks

Pedro Henrique Oliveira Pinheiro

Object recognition is one of the most important problems in computer vision. However, visual recognition poses many challenges when tried to be reproduced by artificial systems. A main challenge is the problem of variability: objects can appear across huge variations in pose, appearance, illumination and occlusion, and a visual system need to be robust to all these changes. In the present thesis, we are interested in pixel-level recognition problems, ie, problems in which the objective is to partition a given image into multiple regions (overlapping or not) that are considered meaningful according to some criterion. Our interests are in algorithms that require the least amount of feature engineering and are easy to scale. Deep learning methods fit very well with this objective: these models alleviate the need of engineered features by discriminatively training a system from raw data (pixels). More precisely, we propose different convolutional neural network (CNN) based algorithms to deal with three important segmentation problems: semantic segmentation, object proposal generation and object detection with segments. The objective of semantic segmentation is to generate a categorical label to each pixel present in a scene. We first study the problem of fully supervised semantic segmentation. We propose a recurrent CNN that is able to consider a large input context (while limiting its capacity), which is essential to model long range pixel label dependencies. This approach achieves state-of-the-art performance without relying on any post-processing smoothing step. However, having densely labeled images to train a model can be expensive and require a lot of human labor. We also propose a CNN-based model that is able to infer object semantic segmentation by leveraging only the object category information from images. This is achieved by casting the problem into a multiple instance learning framework. This approach beats previous state of the art in weakly supervised semantic segmentation by a large margin. Object proposal algorithms generate a set of regions (segments) that are likely to contain objects, independent of their semantic category. Contrary to most approaches (which rely on low-level vision cues), we propose a CNN-based discriminative approach that is able to learn segmentation proposals from raw pixels. This approach is proven to be quite effective in this setting, achieving substantially higher recall using fewer proposals than other methods. The state of the art is pushed further with the introduction of a new top-down network augmentation. The resulting bottom-up/top-down network combines low-level rich spatial information with high-level object semantic information to improve segmentation, while remaining fast at test time. Finally, we show that the proposals generated by our approach, when coupled with a standard state-of-the-art object detection pipeline, achieve considerably better performance than previous proposals methods.

EPFL2017

Lifelong Machine Learning with Data Efficiency and Knowledge Retention

Fei Mi

Artificial intelligence (AI) and machine learning (ML) have become de facto tools in many real-life applications to offer a wide range of benefits for individuals and our society. A classic ML model is typically trained with a large-scale static dataset in an offline manner. Therefore, it can not quickly capture new knowledge in non-stationary environments, and it is difficult to maintain long-term memory for knowledge learned earlier. In practice, many ML systems often need to learn new knowledge (e.g., domains, tasks, distributions, etc.) as more data and experiences are collected, which is referred to as a lifelong ML paradigm in this thesis. We focus on two fundamental challenges to achieve lifelong learning. The first challenge is to quickly learn new knowledge with a small number of observations, and we refer to it as data efficiency. The second challenge is to prevent an ML system from forgetting the old knowledge it has previously learned, and we refer to this challenge as knowledge retention. These two capabilities are crucial for applying ML to most practical applications. In this thesis, we study three important applications with these two challenges, including recommendation systems, task-oriented dialog systems, and the image classification task. First, we propose two approaches to improve data efficiency for task-oriented dialog systems. The first proposed approach is based on Meta-learning, aiming to learn a better model parameter initialization from training data. It can quickly reach a good parameter region of new domains or tasks with a small number of labeled data. The second proposal takes a semi-supervised self-training approach to iteratively train a better model using sufficient unlabeled data when only a limited number of labeled data are available. We empirically demonstrate that both approaches effectively improve data efficiency to learn new knowledge. The second self-training method even consistently improves state-of-the-art large-scale pre-trained models. Second, we tackle the knowledge retention challenge to mitigate the detrimental catastrophic forgetting issue when neural networks learn new knowledge sequentially. We formulate and investigate the ``continual learning'' setting for task-oriented dialog systems and recommendation systems. Through extensive empirical evaluation and analysis, we demonstrate the importance of (1) exemplar replay: storing representative historical data and replaying them to the model while learning new knowledge; (2) dynamic regularization: applying a dynamic regularization term to put flexible constraints on not forgetting previously learned knowledge in each model update cycle. Lastly, we conduct several initial attempts to achieve both data efficiency and knowledge retention in a unified framework. In the recommendation scenario, we propose two approaches using different non-parametric memory modules to retain long-term knowledge. More importantly, the two proposed non-parametric predictions computed on top of them help learn and memorize new knowledge in a data-efficient manner. Apart from the recommendation scenario, we propose a probabilistic evaluation protocol in the widely studied image classification domain. It is general and versatile to simulate a wide range of realistic lifelong learning scenarios that require both knowledge retention and data efficiency for studying different techniques. Through experiments, we also demonstrate the benefit

EPFL2021