The eyes are the windows for humans to interact with the outside world and a prerequisite for the promotion of many applications. In gaze estimation, although the full-face image encompasses the eyes, when extracting gaze features, the network often comprehensively considers the information characteristics of the entire face and its interrelationships. This may lead to neglecting the delicate gaze-related information contained in the eye images. To improve the accuracy of gaze direction prediction, this study proposes a novel multi-stream binocular gaze estimation model named CO-Net, which takes facial and binocular images as input. First, we use facial and eye features to preliminarily estimate a primary gaze direction. Considering the differences that arise when observing objects with both eyes in an unconstrained environment, we utilize this characteristic to purify the gaze features. Finally, we use the purified features to linearly optimize the primary gaze direction. Experimental results demonstrate that our method reduces the mean angular error to 3.63 ∘ on MPIIGaze and 5.9 ∘ on RT-GENE, representing improvements of 7% and 10%, respectively, over the previous state-of-the-art.

Urban time series forecasting is crucial for smart city development and is key to sustainable urban management. Although urban time series models (UTSMs) are effective in general forecasting, they often overlook low-frequency events, such as holidays and extreme weather, leading to degraded performance in practical applications. In this paper, we first investigate how UTSMs handle these infrequent patterns from a neural perspective. Based on our findings, we propose $\textbf{P}$attern $\textbf{N}$euron guided $\textbf{Train}$ing ($\texttt{PN-Train}$), a novel training method that features (i) a $\textit{perturbation-based detector}$ to identify neurons responsible for low-frequency patterns in UTSMs, and (ii) a $\textit{fine-tuning mechanism}$ that enhances these neurons without compromising representation learning on high-frequency patterns. Empirical results demonstrate that $\texttt{PN-Train}$ considerably improves forecasting accuracy for low-frequency events while maintaining high performance for high-frequency events. The code is available at https://github.com/cwang-nus/PN-Train.

The intrinsic non-stationarity of urban spatiotemporal (ST) streams, particularly unique distribution shifts that evolve over time, poses substantial challenges for accurate urban ST forecasting. Existing works often overlook these dynamic shifts, limiting their ability to adapt to evolving trends effectively. To address this challenge, we propose DOL, a novel Distribution-aware Online Learning framework designed to handle the unique shifts in urban ST streams. DOL introduces a streaming update mechanism that leverages streaming memories to strategically adapt to gradual distribution shifts. By aligning network updates with these shifts, DOL avoids unnecessary updates, reducing computational overhead while improving prediction accuracy. DOL also incorporates an adaptive spatiotemporal network with a location-specific learner, enabling it to handle diverse urban distribution shifts across locations. Experimental results on four real-world datasets confirm DOL's superiority over state-of-the-art models. The source code is available at https://github.com/cwang-nus/DOL.

Citywide spatio-temporal (ST) forecasting is a fundamental task for many urban applications, including traffic accident prediction, taxi demand planning, and crowd flow forecasting. The goal of this task is to generate accurate predictions concurrently for all regions within a city. Prior works take great effort on modeling the ST correlations. However, they often overlook intrinsic correlations and inherent data distribution across the city, both of which are influenced by urban zoning and functionality, resulting in inferior performance on citywide ST forecasting. In this paper, we introduce CityCAN, a novel causal attention network, to collectively generate predictions for every region of a city. We first present a causal framework to identify useful correlations among regions, filtering out useless ones, via an intervention strategy. In the framework, a Global Local-Attention Encoder, which leverages attention mechanisms, is designed to jointly learn both local and global ST correlations among correlated regions. Then, we design a citywide loss to constrain the prediction distribution by incorporating the citywide distribution. Extensive experiments on three real-world applications demonstrate the effectiveness of CityCAN.

Traffic management centres worldwide are facing unprecedented challenges from traffic incidents, impacting congestion, delays, pollution, and costs. Congested roads during peak hours are prone to accidents, causing significant disruptions and delays. Non-recurrent congestion such as accidents, occur unexpectedly and are major contributors to delays while disruptions in public transportation lead to increased demand for alternative modes thereby causing severe congestion. Understanding the impact of these incidents on multi-modal traffic networks is crucial for developing effective response plans. This study addresses the limitations of using only historical data to run simulations in advance. The proposed architecture enables optimized simulations with real-time inputs and accurate parameter prediction for improved response planning. This work provides a foundational proof of concept for utilizing minimal real-time data for accurate simulations of dynamic traffic behaviours through a Singapore case study. Future research can focus on comparative analysis methods, optimization strategies, and incorporating pedestrian effects. These advancements will aid transport authorities in making informed decisions for infrastructure and route planning.

The recent expansion in sensor deployment throughout cities has produced extensive traffic data, empowering the traffic management center to closely monitor traffic conditions. These sensors generate massive spatio-temporal (ST) data, serving as the resource that can be harnessed to forecast traffic conditions within the city. Accurate forecasting of these conditions is crucial for traffic management centers, which can facilitate a wide range of urban applications, such as intelligent transportation, traffic simulation, and infrastructure planning. In this summary, we focus on leveraging ST traffic data through AI techniques for forecasting applications. We design novel methodologies to extract knowledge from the traffic data, providing support for a wide variety of real-world applications, including traffic simulation. To achieve it, we introduce the ST forecasting algorithms that operate at two distinct levels: the region-based level and the road-based level. These predictions can be then utilized as input to a traffic simulation framework, enabling the generation of proactive and advantageous planning strategies for urban transportation management. At the region-based level, a city is divided into multiple regions, i.e., even grid cells. Our goal is to generate accurate predictions concurrently for all regions within a city. We first investigate the inherent characteristics of traffic data, i.e., periodicity and ST dependency. We then devise a novel periodic residual learning network (PRNet) to better model the periodicity in traffic ST data. PRNet frames traffic condition forecasting as a periodic residual learning problem by modeling the variation between past and future conditions. Compared to directly predicting traffic conditions that are highly dynamic, learning more stationary variation is much easier, which facilitates model training. Also, the learned variation enables the network to produce residuals between future conditions and its corresponding weekly observations at each time interval and thus contributes to substantially more accurate predictions. At the road-based level, we focus on modeling traffic conditions of individual roads within a region, with the aim of accurately predicting traffic conditions for each road. Among various traffic conditions, traffic incidents are important to traffic management centers, as they can cause significant disruptions and delays. Thus, we introduce an incident forecasting framework, utilizing ST forecasting algorithms, to infer the incident in advance, allowing for more proactive and responsive traffic management strategies. Specifically, we adopt the ST forecasting algorithm to generate each road's speed condition. Considering that speed conditions do not change dramatically within short intervals, we use the predicted speed to infer if a traffic incident has taken place on a road. If there is a significant deviation in the predicted speed conditions from the previous speed behavior, we generate a severity level of the incident based on the deviations. This severity level can be then integrated into simulations, allowing the system to generate response plans in advance.

Crowd flow forecasting, which aims to predict the crowds entering or leaving certain regions, is a fundamental task in smart cities. One of the key properties of crowd flow data is periodicity: a pattern that occurs at regular time intervals, such as a weekly pattern. To capture such periodicity, existing studies either fuse the periodic hidden states into channels for networks to learn or apply extra periodic strategies to the network architecture. In this paper, we devise a novel periodic residual learning network (PRNet) for a better modeling of periodicity in crowd flow data. Unlike existing methods, PRNet frames the crowd flow forecasting as a periodic residual learning problem by modeling the variation between the inputs (the previous time period) and the outputs (the future time period). Compared to directly predicting crowd flows that are highly dynamic, learning more stationary deviation is much easier, which thus facilitates the model training. Besides, the learned variation enables the network to produce the residual between future conditions and its corresponding weekly observations at each time interval, and therefore contributes to substantially more accurate multi-step ahead predictions. Extensive experiments show that PR-Net can be easily integrated into existing models to enhance their predictive performance.

The primary goal of image inpainting is to fix holes in a damaged image with natural contents. A key challenge is that a damaged image contains complex structures in differ-ent ways, with each consisting of its configuration of edges and spatial dependencies. As a result, filled images often converge to unnatural and implausible results. Currently, the edge-image inpainting methods adopt two stages to recover edges and images successively, which suffer from feature in-consistency and error accumulation. This leads us to present a parallel edge-image learning framework that explicitly char-acterizes these internal configurations in a single stage. The framework introduces a dual parallel network-based decoder to generate the image and the edges concurrently, leading to feature consistency at the semantic level. Also, a new cross-fire mechanism aims to exchange edge-image information in the decoder, avoiding error accumulation. Empirical evaluations on benchmark datasets suggest that our approach out-performs the state-of-the-art methods on image inpainting.

Predicting the future paths of an agent's neighbors accurately and in a timely manner is central to the autonomous applications for collision avoidance. Conventional approaches, e.g., LSTM-based models, take considerable computational costs in the prediction, especially for the long sequence prediction. To support more efficient and accurate trajectory predictions, we propose a novel CNN-based spatial-temporal graph framework GraphTCN, which models the spatial interactions as social graphs and captures the spatio-temporal interactions with a modified temporal convolutional network. In contrast to conventional models, both the spatial and temporal modeling of our model are computed within each local time window. Therefore, it can be executed in parallel for much higher efficiency, and meanwhile with accuracy comparable to best-performing approaches. Experimental results confirm that our model achieves better performance in terms of both efficiency and accuracy as compared with state-of-the-art models on various trajectory prediction benchmark datasets.