IIC - Gestion

Institute

Estimating flexibility preferences to resolve temporal scheduling conflicts in activity-based modelling

Michel Bierlaire, Tim Hillel, Janody Pougala

This paper presents a novel activity-based demand model that combines an optimisation framework for continuous temporal scheduling decisions (i.e. activity timings and durations) with traditional discrete choice models for non-temporal choice dimensions (i.e. activity participation, number and type of tours, and destinations). The central idea of our approach is that individuals resolve temporal scheduling conflicts that arise from overlapping activities, e.g. needing to work and desiring to shop at the same time, in order to maximise their daily utility. Flexibility parameters capture behavioural preferences that penalise deviations from desired timings. This framework has three advantages over existing activity-based modelling approaches: (i) the time conflicts between different temporal scheduling decisions including the activity sequence are treated jointly; (ii) flexibility parameters follow a utility maximisation approach; and (iii) the framework can be used to estimate and simulate a city-scale case study in reasonable time. We introduce an estimation routine that allows flexibility parameters to be estimated using real-world observations as well as a simulation routine to efficiently resolve temporal conflicts using an optimisation model. The framework is applied to the full-time workers of a synthetic population for the city of Lausanne, Switzerland. We validate the model results against reported schedules. The results demonstrate the capabilities of our approach to reproduce empirical observations in a real-world case study.

SPRINGER2022

Motion Style Transfer: Modular Low-Rank Adaptation for Deep Motion Forecasting

Alexandre Massoud Alahi, Parth Ashit Kothari, Danya Li, Yuejiang Liu

Deep motion forecasting models have achieved great success when trained on a massive amount of data. Yet, they often perform poorly when training data is limited. To address this challenge, we propose a transfer learning approach for efficiently adapting pre-trained forecasting models to new domains, such as unseen agent types and scene contexts. Unlike the conventional fine-tuning approach that updates the whole encoder, our main idea is to reduce the amount of tunable parameters that can precisely account for the target domain-specific motion style. To this end, we introduce two components that exploit our prior knowledge of motion style shifts: (i) a low-rank motion style adapter that projects and adjusts the style features at a low-dimensional bottleneck; and (ii) a modular adapter strategy that disentangles the features of scene context and motion history to facilitate a fine-grained choice of adaptation layers. Through extensive experimentation, we show that our proposed adapter design, coined MoSA, outperforms prior methods on several forecasting benchmarks.

2022

VL4Pose: Active Learning Through Out-Of-Distribution Detection For Pose Estimation

Alexandre Massoud Alahi

Advances in computing have enabled widespread access to pose estimation, creating new sources of data streams. Unlike mock set-ups for data collection, tapping into these data streams through on-device active learning allows us to directly sample from the real world to improve the spread of the training distribution. However, on-device computing power is limited, implying that any candidate active learning algorithm should have a low compute footprint while also being reliable. Although multiple algorithms cater to pose estimation, they either use extensive compute to power state-of-the-art results or are not competitive in low-resource settings. We address this limitation with VL4Pose (Visual Likelihood For Pose Estimation), a first principles approach for active learning through out-of-distribution detection. We begin with a simple premise: pose estimators often predict incoherent poses for out-of-distribution samples. Hence, can we identify a distribution of poses the model has been trained on, to identify incoherent poses the model is unsure of? Our solution involves modelling the pose through a simple parametric Bayesian network trained via maximum likelihood estimation. Therefore, poses incurring a low likelihood within our framework are out-of-distribution samples making them suitable candidates for annotation. We also observe two useful side-outcomes: VL4Pose in-principle yields better uncertainty estimates by unifying joint and pose level ambiguity, as well as the unintentional but welcome ability of VL4Pose to perform pose refinement in limited scenarios. We perform qualitative and quantitative experiments on three datasets: MPII, LSP and ICVL, spanning human and hand pose estimation. Finally, we note that VL4Pose is simple, computationally inexpensive and competitive, making it suitable for challenging tasks such as on-device active learning.

BMVA Press2022

Composite Relationship Fields with Transformers for Scene Graph Generation

George Adaimi, Alexandre Massoud Alahi, David Mizrahi

Scene graph generation (SGG) methods extract relationships between objects. While most methods focus on improving top-down approaches, which build a scene graph based on detected objects from an off-the-shelf object detector, there is a limited amount of work on bottom-up approaches, which jointly detect objects and their relationships in a single stage. In this work, we present a novel bottom-up SGG approach by representing relationships using Composite Relationship Fields (CoRF). CoRF turns relationship detection into a dense regression and classification task, where each cell of the output feature map identifies surrounding objects and their relationships. Furthermore, we propose a refinement head that leverages Transformers for global scene reasoning, resulting in more meaningful relationship predictions. By combining both contributions, our method outperforms previous bottom-up methods on the Visual Genome dataset by 26% while preserving real-time performance.

2023

Bridging the gap from test rooms to field-tests for human indoor comfort studies: A critical review of the sustainability potential of living laboratories

Mengdi Kong, Dusan Licina

Occupants play a key role in determining final building energy consumption. Empirical evidence must support occupants' modelling. Experiments on human responses to Indoor Environmental Quality (IEQ) are usually performed in test rooms or as in-field monitoring. Between these two approaches, living laboratories, often abbreviated as living labs, represent a valid alternative due to their resemblance to real-world settings. This allows observing actual behaviours while keeping the capability to reliably monitor and control the indoor environment. This work systematically reviewed the available information from 34 living labs for human comfort studies worldwide to define the scope, characteristics, and significance of living labs, for the first time. Most of the reviewed living labs are office environments, and only a few do not involve a university research institution in their operation and management. Most of them are in Europe and the United States, whereas there is a lack of such facilities in other locations and climate zones (e.g., tropics). A larger number of comfort studies in living labs is required to clarify the differences in the knowledge acquired in these experiments compared to in-field and laboratory ones. The review shows that living labs add opportunities for testing and optimizing in-novations in human-centric solutions for comfortable green buildings. Through the living labs approach it is possible to holistically capture the influence of IEQ on occupant perception and the related response, to gather data on larger and more diverse groups of people, and to conduct multi-domain comfort studies involving multidisciplinary approaches given their real-life settings.

ELSEVIER2022

Study on the internal crack network of the ASR-affected concrete by the tomography-based numerical model

Emil Gallyamov, Jean-François Molinari, Mahdieh Shakoorioskooie

In this paper, we present a novel finite element model to simulate alkali-silica reaction in a realistic concrete meso-structure. Application of the internal ASR loading leads to the evolution of multiple deviated cracks and corresponding macroscopic expansion. A particular crack-extension algorithm and a solution scheme provide numerical stability and allow to model complicated crack patterns while preserving the physics. The predictive validity of the model is demonstrated by matching an analytical solution of a loaded penny-shaped crack. The model is applied to the experimental dataset comprising the time-evolving X-ray tomograms of the ASR-affected concrete. Similar to the tomography data, the model results in an expanded concrete sample with a developed crack network. Two hypotheses on the crack loading and extension mechanisms are tested by comparing the crack statistics. Simulations with varying number of ASR sites and application of the uniaxial loading bring interesting insights. The latter concerns the role of the ASR-sites number, the individual loading amplitude and the difference in crack patterns.

PERGAMON-ELSEVIER SCIENCE LTD2022

Vehicle trajectory prediction works, but not everywhere

Alexandre Massoud Alahi, Mohammadhossein Bahari, Seyed Mohsen Moosavi Dezfooli, Saeed Saadatnejad

Vehicle trajectory prediction is nowadays a fundamental pillar of self-driving cars. Both the industry and research communities have acknowledged the need for such a pillar by providing public benchmarks. While state-of-the-art methods are impressive, i.e., they have no off-road prediction, their generalization to cities outside of the benchmark remains unexplored. In this work, we show that those methods do not generalize to new scenes. We present a method that automatically generates realistic scenes causing state-of-the-art models to go off-road. We frame the problem through the lens of adversarial scene generation. The method is a simple yet effective generative model based on atomic scene generation functions along with physical constraints. Our experiments show that more than 60% of existing scenes from the current benchmarks can be modified in a way to make prediction methods fail (i.e., predicting off-road). We further show that the generated scenes (i) are realistic since they do exist in the real world, and (ii) can be used to make existing models more robust, yielding 30-40% reductions in the off-road rate. The code is available online: https://s-attack.github.io/

IEEE2022

Quasi-static crack front deformations in cohesive materials

Mathias Alexandre David Lebihain, Jean-François Molinari, Thibault Didier Roch

When a crack interacts with material heterogeneities, its front distorts and adopts complex tortuous configurations that are reminiscent of the energy barriers encountered during crack propagation. As such, the study of crack front deformations is key to rationalize the effective failure properties of micro-structured solids and interfaces. Yet, the impact of a localized dissipation in a finite region behind the crack front, called the process zone, has often been overlooked. In this work, we derive the equation ruling 3D coplanar crack propagation in heterogeneous cohesive materials where the opening of the crack is resisted by some traction in its wake. We show that the presence of a process zone results in two competing effects on the deformation of crack fronts: (i) it makes the front more compliant to small-wavelength perturbations, and (ii) it smooths out local fluctuations of strength and process zone size, from which emerge heterogeneities of fracture energy. Their respective influence on front deformations is shown to strongly impact the stability of perturbed crack fronts, as well as their stationary shapes when interacting with arrays of tough obstacles. Overall, our theory provides a unified framework to predict the variety of front profiles observed in experiments, even when the small-scale yielding hypothesis of linear elastic fracture mechanics breaks down.

PERGAMON-ELSEVIER SCIENCE LTD2022

TTT++: When Does Self-Supervised Test-Time Training Fail or Thrive?

Alexandre Massoud Alahi, Parth Ashit Kothari, Yuejiang Liu, Taylor Ferdinand Mordan, Bastien Germain C. Van Delft

Test-time training (TTT) through self-supervised learning (SSL) is an emerging paradigm to tackle distributional shifts. Despite encouraging results, it remains unclear when this approach thrives or fails. In this work, we first provide an in-depth look at its limitations and show that TTT can possibly deteriorate, instead of improving, the test-time performance in the presence of severe distribution shifts. To address this issue, we introduce a test-time feature alignment strategy utilizing offline feature summarization and online moment matching, which regularizes adaptation without revisiting training data. We further scale this strategy in the online setting through batch-queue decoupling to enable robust moment estimates even with limited batch size. Given aligned feature distributions, we then shed light on the strong potential of TTT by theoretically analyzing its performance post adaptation. This analysis motivates our use of more informative self-supervision in the form of contrastive learning for visual recognition problems. We empirically demonstrate that our modified version of test-time training, termed TTT++, outperforms state-of-the-art methods by significant margins on several benchmarks. Our result indicates that storing and exploiting extra information, in addition to model parameters, can be a promising direction towards robust test-time adaptation.

2021