S. -H. Gary Chan

Jierun Chen, Shiu-hong Kao, Hao He et al.

To design fast neural networks, many works have been focusing on reducing the number of floating-point operations (FLOPs). We observe that such reduction in FLOPs, however, does not necessarily lead to a similar level of reduction in latency. This mainly stems from inefficiently low floating-point operations per second (FLOPS). To achieve faster networks, we revisit popular operators and demonstrate that such low FLOPS is mainly due to frequent memory access of the operators, especially the depthwise convolution. We hence propose a novel partial convolution (PConv) that extracts spatial features more efficiently, by cutting down redundant computation and memory access simultaneously. Building upon our PConv, we further propose FasterNet, a new family of neural networks, which attains substantially higher running speed than others on a wide range of devices, without compromising on accuracy for various vision tasks. For example, on ImageNet-1k, our tiny FasterNet-T0 is $2.8\times$, $3.3\times$, and $2.4\times$ faster than MobileViT-XXS on GPU, CPU, and ARM processors, respectively, while being $2.9\%$ more accurate. Our large FasterNet-L achieves impressive $83.5\%$ top-1 accuracy, on par with the emerging Swin-B, while having $36\%$ higher inference throughput on GPU, as well as saving $37\%$ compute time on CPU. Code is available at \url{https://github.com/JierunChen/FasterNet}.

Jierun Chen, Tianlang He, Weipeng Zhuo et al.

As convolution has empowered many smart applications, dynamic convolution further equips it with the ability to adapt to diverse inputs. However, the static and dynamic convolutions are either layout-agnostic or computation-heavy, making it inappropriate for layout-specific applications, e.g., face recognition and medical image segmentation. We observe that these applications naturally exhibit the characteristics of large intra-image (spatial) variance and small cross-image variance. This observation motivates our efficient translation variant convolution (TVConv) for layout-aware visual processing. Technically, TVConv is composed of affinity maps and a weight-generating block. While affinity maps depict pixel-paired relationships gracefully, the weight-generating block can be explicitly overparameterized for better training while maintaining efficient inference. Although conceptually simple, TVConv significantly improves the efficiency of the convolution and can be readily plugged into various network architectures. Extensive experiments on face recognition show that TVConv reduces the computational cost by up to 3.1x and improves the corresponding throughput by 2.3x while maintaining a high accuracy compared to the depthwise convolution. Moreover, for the same computation cost, we boost the mean accuracy by up to 4.21%. We also conduct experiments on the optic disc/cup segmentation task and obtain better generalization performance, which helps mitigate the critical data scarcity issue. Code is available at https://github.com/JierunChen/TVConv.

Haoyue Bai, Ceyuan Yang, Yinghao Xu et al.

Deep neural networks achieve superior performance for learning from independent and identically distributed (i.i.d.) data. However, their performance deteriorates significantly when handling out-of-distribution (OoD) data, where the training and test are drawn from different distributions. In this paper, we explore utilizing the generative models as a data augmentation source for improving out-of-distribution robustness of neural classifiers. Specifically, we develop a simple yet effective method called Generative Interpolation to fuse generative models trained from multiple domains for synthesizing diverse OoD samples. Training a generative model directly on the source domains tends to suffer from mode collapse and sometimes amplifies the data bias. Instead, we first train a StyleGAN model on one source domain and then fine-tune it on the other domains, resulting in many correlated generators where their model parameters have the same initialization thus are aligned. We then linearly interpolate the model parameters of the generators to spawn new sets of generators. Such interpolated generators are used as an extra data augmentation source to train the classifiers. The interpolation coefficients can flexibly control the augmentation direction and strength. In addition, a style-mixing mechanism is applied to further improve the diversity of the generated OoD samples. Our experiments show that the proposed method explicitly increases the diversity of training domains and achieves consistent improvements over baselines across datasets and multiple different distribution shifts.

Weipeng Zhuo, Ka Ho Chiu, Jierun Chen et al.

Floor labels of crowdsourced RF signals are crucial for many smart-city applications, such as multi-floor indoor localization, geofencing, and robot surveillance. To build a prediction model to identify the floor number of a new RF signal upon its measurement, conventional approaches using the crowdsourced RF signals assume that at least few labeled signal samples are available on each floor. In this work, we push the envelope further and demonstrate that it is technically feasible to enable such floor identification with only one floor-labeled signal sample on the bottom floor while having the rest of signal samples unlabeled. We propose FIS-ONE, a novel floor identification system with only one labeled sample. FIS-ONE consists of two steps, namely signal clustering and cluster indexing. We first build a bipartite graph to model the RF signal samples and obtain a latent representation of each node (each signal sample) using our attention-based graph neural network model so that the RF signal samples can be clustered more accurately. Then, we tackle the problem of indexing the clusters with proper floor labels, by leveraging the observation that signals from an access point can be detected on different floors, i.e., signal spillover. Specifically, we formulate a cluster indexing problem as a combinatorial optimization problem and show that it is equivalent to solving a traveling salesman problem, whose (near-)optimal solution can be found efficiently. We have implemented FIS-ONE and validated its effectiveness on the Microsoft dataset and in three large shopping malls. Our results show that FIS-ONE outperforms other baseline algorithms significantly, with up to 23% improvement in adjusted rand index and 25% improvement in normalized mutual information using only one floor-labeled signal sample.

Shuhan Zhong, Sizhe Song, Weipeng Zhuo et al.

Time series data, including univariate and multivariate ones, are characterized by unique composition and complex multi-scale temporal variations. They often require special consideration of decomposition and multi-scale modeling to analyze. Existing deep learning methods on this best fit to univariate time series only, and have not sufficiently considered sub-series modeling and decomposition completeness. To address these challenges, we propose MSD-Mixer, a Multi-Scale Decomposition MLP-Mixer, which learns to explicitly decompose and represent the input time series in its different layers. To handle the multi-scale temporal patterns and multivariate dependencies, we propose a novel temporal patching approach to model the time series as multi-scale patches, and employ MLPs to capture intra- and inter-patch variations and channel-wise correlations. In addition, we propose a novel loss function to constrain both the mean and the autocorrelation of the decomposition residual for better decomposition completeness. Through extensive experiments on various real-world datasets for five common time series analysis tasks, we demonstrate that MSD-Mixer consistently and significantly outperforms other state-of-the-art algorithms with better efficiency.

Zhe Li, Cheng Chi, Boan Zhu et al.

Humans learn locomotion through visual observation, interpreting visual content first before imitating actions. However, state-of-the-art humanoid locomotion systems rely on either curated motion capture trajectories or sparse text commands, leaving a critical gap between visual understanding and control. Text-to-motion methods suffer from semantic sparsity and staged pipeline errors, while video-based approaches only perform mechanical pose mimicry without genuine visual understanding. We propose RoboMirror, the first retargeting-free video-to-locomotion framework embodying "understand before you imitate". Leveraging VLMs, it distills raw egocentric/third-person videos into visual motion intents, which directly condition a diffusion-based policy to generate physically plausible, semantically aligned locomotion without explicit pose reconstruction or retargeting. Extensive experiments validate the effectiveness of RoboMirror, it enables telepresence via egocentric videos, drastically reduces third-person control latency by 80%, and achieves a 3.7% higher task success rate than baselines. By reframing humanoid control around video understanding, we bridge the visual understanding and action gap.

Tianlang He, Zhongming Lin, Tianrui Jiang et al.

The Long-Range (LoRa) protocol, known for its extensive range and low power, has increasingly been adopted in tags worn by mentally incapacitated persons (MIPs) and others at risk of going missing. We study the sequential decision-making process for a mobile sensor to locate a periodically broadcasting LoRa tag with the fewest moves (hops) in general, unknown environments, guided by the received signal strength indicator (RSSI). While existing methods leverage reinforcement learning for search, they remain vulnerable to domain shift and signal fluctuation, resulting in cascading decision errors that culminate in substantial localization inaccuracies. To bridge this gap, we propose LoRaCompass, a reinforcement learning model designed to achieve robust and efficient search for a LoRa tag. For exploitation under domain shift and signal fluctuation, LoRaCompass learns a robust spatial representation from RSSI to maximize the probability of moving closer to a tag, via a spatially-aware feature extractor and a policy distillation loss function. It further introduces an exploration function inspired by the upper confidence bound (UCB) that guides the sensor toward the tag with increasing confidence. We have validated LoRaCompass in ground-based and drone-assisted scenarios within diverse unseen environments covering an area of over 80km^2. It has demonstrated high success rate (>90%) in locating the tag within 100m proximity (a 40% improvement over existing methods) and high efficiency with a search path length (in hops) that scales linearly with the initial distance.

Jierun Chen, Fangyun Wei, Jinjing Zhao et al.

Referring expression comprehension (REC) involves localizing a target instance based on a textual description. Recent advancements in REC have been driven by large multimodal models (LMMs) like CogVLM, which achieved 92.44% accuracy on RefCOCO. However, this study questions whether existing benchmarks such as RefCOCO, RefCOCO+, and RefCOCOg, capture LMMs' comprehensive capabilities. We begin with a manual examination of these benchmarks, revealing high labeling error rates: 14% in RefCOCO, 24% in RefCOCO+, and 5% in RefCOCOg, which undermines the authenticity of evaluations. We address this by excluding problematic instances and reevaluating several LMMs capable of handling the REC task, showing significant accuracy improvements, thus highlighting the impact of benchmark noise. In response, we introduce Ref-L4, a comprehensive REC benchmark, specifically designed to evaluate modern REC models. Ref-L4 is distinguished by four key features: 1) a substantial sample size with 45,341 annotations; 2) a diverse range of object categories with 365 distinct types and varying instance scales from 30 to 3,767; 3) lengthy referring expressions averaging 24.2 words; and 4) an extensive vocabulary comprising 22,813 unique words. We evaluate a total of 24 large models on Ref-L4 and provide valuable insights. The cleaned versions of RefCOCO, RefCOCO+, and RefCOCOg, as well as our Ref-L4 benchmark and evaluation code, are available at https://github.com/JierunChen/Ref-L4.

Zhuoxuan Peng, S. -H. Gary Chan

Due to its promising results, density map regression has been widely employed for image-based crowd counting. The approach, however, often suffers from severe performance degradation when tested on data from unseen scenarios, the so-called "domain shift" problem. To address the problem, we investigate in this work single domain generalization (SDG) for crowd counting. The existing SDG approaches are mainly for image classification and segmentation, and can hardly be extended to our case due to its regression nature and label ambiguity (i.e., ambiguous pixel-level ground truths). We propose MPCount, a novel effective SDG approach even for narrow source distribution. MPCount stores diverse density values for density map regression and reconstructs domain-invariant features by means of only one memory bank, a content error mask and attention consistency loss. By partitioning the image into grids, it employs patch-wise classification as an auxiliary task to mitigate label ambiguity. Through extensive experiments on different datasets, MPCount is shown to significantly improve counting accuracy compared to the state of the art under diverse scenarios unobserved in the training data characterized by narrow source distribution. Code is available at https://github.com/Shimmer93/MPCount.

Zhuoxuan Peng, Boan Zhu, Xingjian Zhang et al.

Current mmWave datasets for human pose estimation (HPE) are scarce and lack diversity in both point cloud (PC) attributes and human poses, severely hampering the generalization ability of their trained models. On the other hand, unlabeled mmWave HPE data and diverse LiDAR HPE datasets are readily available. We propose EMDUL, a novel approach to expand the volume and diversity of an existing mmWave dataset using unlabeled mmWave data and a LiDAR dataset. EMDUL trains a pseudo-label estimator to annotate the unlabeled mmWave data and is able to convert, or translate, a given annotated LiDAR PC to its mmWave counterpart. Expanded with both LiDAR-converted and pseudo-labeled mmWave PCs, our mmWave dataset significantly boosts the performance and generalization ability of all our HPE models, with substantial 15.1% and 18.9% error reductions for in-domain and out-of-domain settings, respectively.

Dongting Hu, Jierun Chen, Xijie Huang et al.

Existing text-to-image (T2I) diffusion models face several limitations, including large model sizes, slow runtime, and low-quality generation on mobile devices. This paper aims to address all of these challenges by developing an extremely small and fast T2I model that generates high-resolution and high-quality images on mobile platforms. We propose several techniques to achieve this goal. First, we systematically examine the design choices of the network architecture to reduce model parameters and latency, while ensuring high-quality generation. Second, to further improve generation quality, we employ cross-architecture knowledge distillation from a much larger model, using a multi-level approach to guide the training of our model from scratch. Third, we enable a few-step generation by integrating adversarial guidance with knowledge distillation. For the first time, our model SnapGen, demonstrates the generation of 1024x1024 px images on a mobile device around 1.4 seconds. On ImageNet-1K, our model, with only 372M parameters, achieves an FID of 2.06 for 256x256 px generation. On T2I benchmarks (i.e., GenEval and DPG-Bench), our model with merely 379M parameters, surpasses large-scale models with billions of parameters at a significantly smaller size (e.g., 7x smaller than SDXL, 14x smaller than IF-XL).

Zhuoxuan Peng, Yiyi Ding, Yang Lin et al.

Many modern video-based human action recognition (HAR) approaches use 2D skeleton as the intermediate representation in their prediction pipelines. Despite overall encouraging results, these approaches still struggle in many common scenes, mainly because the skeleton does not capture critical action-related information pertaining to the depth of the joints, contour of the human body, and interaction between the human and objects. To address this, we propose an effective approach to augment skeleton with a representation capturing action-related information in the pipeline of HAR. The representation, termed Scale-Body-Flow (SBF), consists of three distinct components, namely a scale map volume given by the scale (and hence depth information) of each joint, a body map outlining the human subject, and a flow map indicating human-object interaction given by pixel-wise optical flow values. To predict SBF, we further present SFSNet, a novel segmentation network supervised by the skeleton and optical flow without extra annotation overhead beyond the existing skeleton extraction. Extensive experiments across different datasets demonstrate that our pipeline based on SBF and SFSNet achieves significantly higher HAR accuracy with similar compactness and efficiency as compared with the state-of-the-art skeleton-only approaches.

Shuhan Zhong, Weipeng Zhuo, Sizhe Song et al.

Irregular multivariate time series (IMTS) is characterized by the lack of synchronized observations across its different channels. In this paper, we point out that this channel-wise asynchrony can lead to poor channel-wise modeling of existing deep learning methods. To overcome this limitation, we propose MTM, a multi-scale token mixing transformer for the classification of IMTS. We find that the channel-wise asynchrony can be alleviated by down-sampling the time series to coarser timescales, and propose to incorporate a masked concat pooling in MTM that gradually down-samples IMTS to enhance the channel-wise attention modules. Meanwhile, we propose a novel channel-wise token mixing mechanism which proactively chooses important tokens from one channel and mixes them with other channels, to further boost the channel-wise learning of our model. Through extensive experiments on real-world datasets and comparison with state-of-the-art methods, we demonstrate that MTM consistently achieves the best performance on all the benchmarks, with improvements of up to 3.8% in AUPRC for classification.

Zhongyi Yu, Zhenghao Wu, Shuhan Zhong et al.

Missing values are a common problem that poses significant challenges to data analysis and machine learning. This problem necessitates the development of an effective imputation method to fill in the missing values accurately, thereby enhancing the overall quality and utility of the datasets. Existing imputation methods, however, fall short of explicitly considering the `missingness' information in the data during the embedding initialization stage and modeling the entangled feature and sample correlations during the learning process, thus leading to inferior performance. We propose M$^3$-Impute, which aims to explicitly leverage the missingness information and such correlations with novel masking schemes. M$^3$-Impute first models the data as a bipartite graph and uses a graph neural network to learn node embeddings, where the refined embedding initialization process directly incorporates the missingness information. They are then optimized through M$^3$-Impute's novel feature correlation unit (FRU) and sample correlation unit (SRU) that effectively captures feature and sample correlations for imputation. Experiment results on 25 benchmark datasets under three different missingness settings show the effectiveness of M$^3$-Impute by achieving 20 best and 4 second-best MAE scores on average.

Tianlang He, Zhiqiu Xia, S. -H. Gary Chan

Knowing a pedestrian's conveyor state of ''elevator,'' ''escalator,'' or ''neither'' is fundamental to many applications such as indoor navigation and people flow management. Previous studies on classifying the conveyor state often rely on specially designed body-worn sensors or make strong assumptions on pedestrian behaviors, which greatly strangles their deployability. To overcome this, we study the classification problem under arbitrary pedestrian behaviors using the inertial navigation system (INS) of the commonly available smartphones (including accelerometer, gyroscope, and magnetometer). This problem is challenging, because the INS signals of the conveyor states are entangled by the arbitrary and diverse pedestrian behaviors. We propose ELESON, a novel and lightweight deep-learning approach that uses phone INS to classify a pedestrian to elevator, escalator, or neither. Using causal decomposition and adversarial learning, ELESON extracts the motion and magnetic features of conveyor state independent of pedestrian behavior, based on which it estimates the state confidence by means of an evidential classifier. We curate a large and diverse dataset with 36,420 instances of pedestrians randomly taking elevators and escalators under arbitrary unknown behaviors. Our extensive experiments show that ELESON is robust against pedestrian behavior, achieving a high accuracy of over 0.9 in F1 score, strong confidence discriminability of 0.81 in AUROC (Area Under the Receiver Operating Characteristics), and low computational and memory requirements fit for common smartphone deployment.

Tianlang He, Jiajie Tan, Weipeng Zhuo et al.

Proximity detection is to determine whether an IoT receiver is within a certain distance from a signal transmitter. Due to its low cost and high popularity, Bluetooth low energy (BLE) has been used to detect proximity based on the received signal strength indicator (RSSI). To address the fact that RSSI can be markedly influenced by device carriage states, previous works have incorporated RSSI with inertial measurement unit (IMU) using deep learning. However, they have not sufficiently accounted for the impact of multipath. Furthermore, due to the special setup, the IMU data collected in the training process may be biased, which hampers the system's robustness and generalizability. This issue has not been studied before. We propose PRID, an IMU-assisted BLE proximity detection approach robust against RSSI fluctuation and IMU data bias. PRID histogramizes RSSI to extract multipath features and uses carriage state regularization to mitigate overfitting due to IMU data bias. We further propose PRID-lite based on a binarized neural network to substantially cut memory requirements for resource-constrained devices. We have conducted extensive experiments under different multipath environments, data bias levels, and a crowdsourced dataset. Our results show that PRID significantly reduces false detection cases compared with the existing arts (by over 50%). PRID-lite further reduces over 90% PRID model size and extends 60% battery life, with a minor compromise in accuracy (7%).

Guanyao Li, Shuhan Zhong, S. -H. Gary Chan et al.

We study the forecasting problem for traffic with dynamic, possibly periodical, and joint spatial-temporal dependency between regions. Given the aggregated inflow and outflow traffic of regions in a city from time slots 0 to t-1, we predict the traffic at time t at any region. Prior arts in the area often consider the spatial and temporal dependencies in a decoupled manner or are rather computationally intensive in training with a large number of hyper-parameters to tune. We propose ST-TIS, a novel, lightweight, and accurate Spatial-Temporal Transformer with information fusion and region sampling for traffic forecasting. ST-TIS extends the canonical Transformer with information fusion and region sampling. The information fusion module captures the complex spatial-temporal dependency between regions. The region sampling module is to improve the efficiency and prediction accuracy, cutting the computation complexity for dependency learning from $O(n^2)$ to $O(n\sqrt{n})$, where n is the number of regions. With far fewer parameters than state-of-the-art models, the offline training of our model is significantly faster in terms of tuning and computation (with a reduction of up to $90\%$ on training time and network parameters). Notwithstanding such training efficiency, extensive experiments show that ST-TIS is substantially more accurate in online prediction than state-of-the-art approaches (with an average improvement of up to $9.5\%$ on RMSE, and $12.4\%$ on MAPE).

Haoyue Bai, Fengwei Zhou, Lanqing Hong et al.

Recent advances on Out-of-Distribution (OoD) generalization reveal the robustness of deep learning models against distribution shifts. However, existing works focus on OoD algorithms, such as invariant risk minimization, domain generalization, or stable learning, without considering the influence of deep model architectures on OoD generalization, which may lead to sub-optimal performance. Neural Architecture Search (NAS) methods search for architecture based on its performance on the training data, which may result in poor generalization for OoD tasks. In this work, we propose robust Neural Architecture Search for OoD generalization (NAS-OoD), which optimizes the architecture with respect to its performance on generated OoD data by gradient descent. Specifically, a data generator is learned to synthesize OoD data by maximizing losses computed by different neural architectures, while the goal for architecture search is to find the optimal architecture parameters that minimize the synthetic OoD data losses. The data generator and the neural architecture are jointly optimized in an end-to-end manner, and the minimax training process effectively discovers robust architectures that generalize well for different distribution shifts. Extensive experimental results show that NAS-OoD achieves superior performance on various OoD generalization benchmarks with deep models having a much fewer number of parameters. In addition, on a real industry dataset, the proposed NAS-OoD method reduces the error rate by more than 70% compared with the state-of-the-art method, demonstrating the proposed method's practicality for real applications.

Haoyue Bai, Song Wen, S. -H. Gary Chan

Labeled crowd scene images are expensive and scarce. To significantly reduce the requirement of the labeled images, we propose ColorCount, a novel CNN-based approach by combining self-supervised transfer colorization learning and global prior classification to leverage the abundantly available unlabeled data. The self-supervised colorization branch learns the semantics and surface texture of the image by using its color components as pseudo labels. The classification branch extracts global group priors by learning correlations among image clusters. Their fused resultant discriminative features (global priors, semantics and textures) provide ample priors for counting, hence significantly reducing the requirement of labeled images. We conduct extensive experiments on four challenging benchmarks. ColorCount achieves much better performance as compared with other unsupervised approaches. Its performance is close to the supervised baseline with substantially less labeled data (10\% of the original one).

Haoyue Bai, S. -H. Gary Chan

We study video crowd counting, which is to estimate the number of objects (people in this paper) in all the frames of a video sequence. Previous work on crowd counting is mostly on still images. There has been little work on how to properly extract and take advantage of the spatial-temporal correlation between neighboring frames in both short and long ranges to achieve high estimation accuracy for a video sequence. In this work, we propose Monet, a novel and highly accurate motion-guided non-local spatial-temporal network for video crowd counting. Monet first takes people flow (motion information) as guidance to coarsely segment the regions of pixels where a person may be. Given these regions, Monet then uses a non-local spatial-temporal network to extract spatial-temporally both short and long-range contextual information. The whole network is finally trained end-to-end with a fused loss to generate a high-quality density map. Noting the scarcity and low quality (in terms of resolution and scene diversity) of the publicly available video crowd datasets, we have collected and built a large-scale video crowd counting datasets, VidCrowd, to contribute to the community. VidCrowd contains 9,000 frames of high resolution (2560 x 1440), with 1,150,239 head annotations captured in different scenes, crowd density and lighting in two cities. We have conducted extensive experiments on the challenging VideoCrowd and two public video crowd counting datasets: UCSD and Mall. Our approach achieves substantially better performance in terms of MAE and MSE as compared with other state-of-the-art approaches.

Jierun Chen, Song Wen, S. -H. Gary Chan

Image demosaicking and denoising are the two key fundamental steps in digital camera pipelines, aiming to reconstruct clean color images from noisy luminance readings. In this paper, we propose and study Wild-JDD, a novel learning framework for joint demosaicking and denoising in the wild. In contrast to previous works which generally assume the ground truth of training data is a perfect reflection of the reality, we consider here the more common imperfect case of ground truth uncertainty in the wild. We first illustrate its manifestation as various kinds of artifacts including zipper effect, color moire and residual noise. Then we formulate a two-stage data degradation process to capture such ground truth uncertainty, where a conjugate prior distribution is imposed upon a base distribution. After that, we derive an evidence lower bound (ELBO) loss to train a neural network that approximates the parameters of the conjugate prior distribution conditioned on the degraded input. Finally, to further enhance the performance for out-of-distribution input, we design a simple but effective fine-tuning strategy by taking the input as a weakly informative prior. Taking into account ground truth uncertainty, Wild-JDD enjoys good interpretability during optimization. Extensive experiments validate that it outperforms state-of-the-art schemes on joint demosaicking and denoising tasks on both synthetic and realistic raw datasets.

Haoyue Bai, Jiageng Mao, S. -H. Gary Chan

Single image crowd counting is a challenging computer vision problem with wide applications in public safety, city planning, traffic management, etc. With the recent development of deep learning techniques, crowd counting has aroused much attention and achieved great success in recent years. This survey is to provide a comprehensive summary of recent advances on deep learning-based crowd counting techniques via density map estimation by systematically reviewing and summarizing more than 200 works in the area since 2015. Our goals are to provide an up-to-date review of recent approaches, and educate new researchers in this field the design principles and trade-offs. After presenting publicly available datasets and evaluation metrics, we review the recent advances with detailed comparisons on three major design modules for crowd counting: deep neural network designs, loss functions, and supervisory signals. We study and compare the approaches using the public datasets and evaluation metrics. We conclude the survey with some future directions.

Haoyue Bai, Rui Sun, Lanqing Hong et al.

While deep learning demonstrates its strong ability to handle independent and identically distributed (IID) data, it often suffers from out-of-distribution (OoD) generalization, where the test data come from another distribution (w.r.t. the training one). Designing a general OoD generalization framework to a wide range of applications is challenging, mainly due to possible correlation shift and diversity shift in the real world. Most of the previous approaches can only solve one specific distribution shift, such as shift across domains or the extrapolation of correlation. To address that, we propose DecAug, a novel decomposed feature representation and semantic augmentation approach for OoD generalization. DecAug disentangles the category-related and context-related features. Category-related features contain causal information of the target object, while context-related features describe the attributes, styles, backgrounds, or scenes, causing distribution shifts between training and test data. The decomposition is achieved by orthogonalizing the two gradients (w.r.t. intermediate features) of losses for predicting category and context labels. Furthermore, we perform gradient-based augmentation on context-related features to improve the robustness of the learned representations. Experimental results show that DecAug outperforms other state-of-the-art methods on various OoD datasets, which is among the very few methods that can deal with different types of OoD generalization challenges.

Guanyao Li, Siyan Hu, Shuhan Zhong et al.

Covid-19 is primarily spread through contact with the virus which may survive on surfaces with lifespan of more than hours. To curb its spread, it is hence of vital importance to detect and quarantine those who have been in contact with the virus for sustained period of time, the so-called close contacts. In this work, we study, for the first time, automatic contact detection when the virus has a lifespan. Leveraging upon the ubiquity of WiFi signals, we propose a novel, private, and fully distributed WiFi-based approach called vContact. Users installing an app continuously scan WiFi and store its hashed IDs. Given a confirmed case, the signals of the major places he/she visited are then uploaded to a server and matched with the stored signals of users to detect contact. vContact is not based on phone pairing, and no information of any other users is stored locally. The confirmed case does not need to have installed the app for it to work properly. As WiFi data are sampled sporadically, we propose efficient signal processing approaches and similarity metric to align and match signals of any time. We conduct extensive indoor and outdoor experiments to evaluate the performance of vContact. Our results demonstrate that vContact is efficient and robust for contact detection. The precision and recall of contact detection are high (in the range of 50-90%) for close contact proximity (2m). Its performance is robust with respect to signal lengths (AP numbers) and phone heterogeneity. By implementing vContact as an app, we present a case study to demonstrate the validity of our design in notifying its users their exposure to virus with lifespan.

Haoyue Bai, Song Wen, S. -H. Gary Chan

Crowd counting is to estimate the number of objects (e.g., people or vehicles) in an image of unconstrained congested scenes. Designing a general crowd counting algorithm applicable to a wide range of crowd images is challenging, mainly due to the possibly large variation in object scales and the presence of many isolated small clusters. Previous approaches based on convolution operations with multi-branch architecture are effective for only some narrow bands of scales and have not captured the long-range contextual relationship due to isolated clustering. To address that, we propose SACANet, a novel scale-adaptive long-range context-aware network for crowd counting. SACANet consists of three major modules: the pyramid contextual module which extracts long-range contextual information and enlarges the receptive field, a scale-adaptive self-attention multi-branch module to attain high scale sensitivity and detection accuracy of isolated clusters, and a hierarchical fusion module to fuse multi-level self-attention features. With group normalization, SACANet achieves better optimality in the training process. We have conducted extensive experiments using the VisDrone2019 People dataset, the VisDrone2019 Vehicle dataset, and some other challenging benchmarks. As compared with the state-of-the-art methods, SACANet is shown to be effective, especially for extremely crowded conditions with diverse scales and scattered clusters, and achieves much lower MAE as compared with baselines.

Yifeng Zhang, Ka-Ho Chow, S. -H. Gary Chan

Much sequential data exhibits highly non-uniform information distribution. This cannot be correctly modeled by traditional Long Short-Term Memory (LSTM). To address that, recent works have extended LSTM by adding more activations between adjacent inputs. However, the approaches often use a fixed depth, which is at the step of the most information content. This one-size-fits-all worst-case approach is not satisfactory, because when little information is distributed to some steps, shallow structures can achieve faster convergence and consume less computation resource. In this paper, we develop a Depth-Adaptive Long Short-Term Memory (DA-LSTM) architecture, which can dynamically adjust the structure depending on information distribution without prior knowledge. Experimental results on real-world datasets show that DA-LSTM costs much less computation resource and substantially reduce convergence time by $41.78\%$ and $46.01 \%$, compared with Stacked LSTM and Deep Transition LSTM, respectively.

Ka-Ho Chow, Anish Hiranandani, Yifeng Zhang et al.

Representation learning of pedestrian trajectories transforms variable-length timestamp-coordinate tuples of a trajectory into a fixed-length vector representation that summarizes spatiotemporal characteristics. It is a crucial technique to connect feature-based data mining with trajectory data. Trajectory representation is a challenging problem, because both environmental constraints (e.g., wall partitions) and temporal user dynamics should be meticulously considered and accounted for. Furthermore, traditional sequence-to-sequence autoencoders using maximum log-likelihood often require dataset covering all the possible spatiotemporal characteristics to perform well. This is infeasible or impractical in reality. We propose TREP, a practical pedestrian trajectory representation learning algorithm which captures the environmental constraints and the pedestrian dynamics without the need of any training dataset. By formulating a sequence-to-sequence autoencoder with a spatial-aware objective function under the paradigm of actor-critic reinforcement learning, TREP intelligently encodes spatiotemporal characteristics of trajectories with the capability of handling diverse trajectory patterns. Extensive experiments on both synthetic and real datasets validate the high fidelity of TREP to represent trajectories.