Position Embedding Research Articles

TransCrack: revisiting fine-grained road crack detection with a transformer design.

Prior convolution-based road crack detectors typically learn more abstract visual representation with increasing receptive field via an encoder-decoder architecture. Despite the promising accuracy, progressive spatial resolution reduction causes semantic feature blurring, leading to coarse and incontiguous distress detection. To these ends, an alternative sequence-to-sequence perspective with atransformer network termed TransCrack is introduced for road crack detection. Specifically, an image is decomposed into a grid of fixed-size crack patches, which is flattened with position embedding into a sequence. We further propose a pure transformer-based encoder with multi-head reduced self-attention modules and feed-forward networks for explicitly modelling long-range dependencies from the sequential input in a global receptive field. More importantly, a simple decoder with cross-layer aggregation architecture is developed to incorporate global with local attentions across different regions for detailed feature recovery and pixel-wise crack mask prediction. Empirical studies are conducted on three publicly available damage detection benchmarks. The proposed TransCrack achieves a state-of-the-art performance over all counterparts by a substantialmargin, and qualitative results further demonstrate its superiority in contiguous crack recognition and fine-grained profile extraction. This article is part of the theme issue 'Artificial intelligence in failure analysis of transportation infrastructure and materials'.

A bio-inspired positional embedding network for transformer-based models

Owing to the progress of transformer-based networks, there have been significant improvements in the performance of vision models in recent years. However, there is further potential for improvement in positional embeddings that play a crucial role in distinguishing information across different positions. Based on the biological mechanisms of human visual pathways, we propose a positional embedding network that adaptively captures position information by modeling the dorsal pathway, which is responsible for spatial perception in human vision. Our proposed double-stream architecture leverages large zero-padding convolutions to learn local positional features and utilizes transformers to learn global features, effectively capturing the interaction between dorsal and ventral pathways. To evaluate the effectiveness of our method, we implemented experiments on various datasets, employing differentiated designs. Our statistical analysis demonstrates that the simple implementation significantly enhances image classification performance, and the observed trends demonstrate its biological plausibility.

Position‐aware spatio‐temporal graph convolutional networks for skeleton‐based action recognition

AbstractGraph Convolutional Networks (GCNs) have been widely used in skeleton‐based action recognition. Though significant performance has been achieved, it is still challenging to effectively model the complex dynamics of skeleton sequences. A novel position‐aware spatio‐temporal GCN for skeleton‐based action recognition is proposed, where the positional encoding is investigated to enhance the capacity of typical baselines for comprehending the dynamic characteristics of action sequence. Specifically, the authors’ method systematically investigates the temporal position encoding and spatial position embedding, in favour of explicitly capturing the sequence ordering information and the identity information of nodes that are used in graphs. Additionally, to alleviate the redundancy and over‐smoothing problems of typical GCNs, the authors’ method further investigates a subgraph mask, which gears to mine the prominent subgraph patterns over the underlying graph, letting the model be robust against the impaction of some irrelevant joints. Extensive experiments on three large‐scale datasets demonstrate that our model can achieve competitive results comparing to the previous state‐of‐art methods.

Multi-scale graph neural network for global stereo matching

Currently, deep learning-based stereo matching is solely based on local convolution networks, which lack enough global information for accurate disparity estimation. Motivated by the excellent global representation of the graph, a novel Multi-scale Graph Neural Network (MGNN) is proposed to essentially improve stereo matching from the global aspect. Firstly, we construct the multi-scale graph structure, where the multi-scale nodes with projected multi-scale image features can be directly linked by the inner-scale and cross-scale edges, instead of solely relying on local convolutions for deep learning-based stereo matching. To enhance the spatial position information at non-Euclidean multi-scale graph space, we further propose a multi-scale position embedding to embed the potential position features of Euclidean space into projected multi-scale image features. Secondly, we propose the multi-scale graph feature inference to extract global context information on multi-scale graph structure. Thus, the features not only be globally inferred on each scale, but also can be interactively inferred across different scales to comprehensively consider global context information with multi-scale receptive fields. Finally, MGNN is deployed into dense stereo matching and experiments demonstrate that our method achieves state-of-the-art performance on Scene Flow, KITTI 2012/2015, and Middlebury Stereo Evaluation v.3/2021.

SWSA transformer: A forecasting method of ultra-short-term wind speed from an offshore wind farm using global attention mechanism

Accurate ultra-short-term wind speed forecasting is great significance to ensure large scale integration of wind power into the power grid, but the randomness, instability, and non-linear nature of wind speed make it very difficult to be predicted accurately. To solve this problem, shifted window stationary attention transformer (SWSA transformer) is proposed based on a global attention mechanism for ultra-short-term forecasting of wind speed. SWSA transformer can sufficiently extract these complicated features of wind speed to improve the prediction accuracy of wind speed. First, positional embedding and temporal embedding are added at the bottom of the proposed method structure to mark wind speed series, which enables complicated global features of wind speed to be more effectively extracted by attention. Second, a shifted window is utilized to enhance the ability of attention to capture features from the edge sequences. Third, a stationary attention mechanism is applied to not only extract features of wind speed but also optimize the encoder-decoder network for smoothing wind speed sequences. Finally, the predicted values of wind speed are obtained using the calculation in the decoder network. To verify the proposed method, tests are performed utilizing data from an real offshore wind farm. The results show that the proposed method outperforms many popular models evaluated by many indexes including gated recurrent unit, Gaussian process regression, long-short term memory, shared weight long short-term memory network, and shared weight long short-term memory network -Gaussian process regression, in terms of mean absolute error, mean square error (MSE), root mean square error, mean absolute percentage error, mean square percentage error, and coefficient of determination (R2).

ITran: A novel transformer-based approach for industrial anomaly detection and localization

Anomaly detection is currently an essential quality monitoring process in industrial production. It is often affected by factors such as under or over reconstruction of images and unclear criteria for feature distribution evaluation, thus making it challenging to improve detection performance. To solve the above problems, this paper proposes a novel transformer-based approach, Inductive Transformer (ITran) for industrial anomaly detection and localization, which utilizes a multi-layer pyramid structure and multi-level jump connections to extract multi-scale features of the data, putting the anomaly detection into the feature space and achieving more accurate industrial anomaly detection and localization results. It incorporates inductive bias and convolution operations into the Transformer which helps to break the myth of Transformer being “data hungry”. Compared with the common Transformers, ITran significantly reduces the computational cost and memory usage and makes it work well on small datasets. In addition, we basically eliminate the effect of positional embedding on the proposed Transformer model. Sufficient experiments have been conducted to validate global anomaly detection on three datasets MNIST, Fashion-MNST and Cifar-10, as well as local anomaly detection on the industrial datasets MVTec AD, Concrete Crack Image and BTAD. The proposed ITran achieves outstanding results on all the above datasets.

A Learnable Radial Basis Positional Embedding for Coordinate-MLPs

We propose a novel method to enhance the performance of coordinate-MLPs (also referred to as neural fields) by learning instance-specific positional embeddings. End-to-end optimization of positional embedding parameters along with network weights leads to poor generalization performance. Instead, we develop a generic framework to learn the positional embedding based on the classic graph-Laplacian regularization, which can implicitly balance the trade-off between memorization and generalization. This framework is then used to propose a novel positional embedding scheme, where the hyperparameters are learned per coordinate (i.e instance) to deliver optimal performance. We show that the proposed embedding achieves better performance with higher stability compared to the well-established random Fourier features (RFF). Further, we demonstrate that the proposed embedding scheme yields stable gradients, enabling seamless integration into deep architectures as intermediate layers.

Pose-Oriented Transformer with Uncertainty-Guided Refinement for 2D-to-3D Human Pose Estimation

There has been a recent surge of interest in introducing transformers to 3D human pose estimation (HPE) due to their powerful capabilities in modeling long-term dependencies. However, existing transformer-based methods treat body joints as equally important inputs and ignore the prior knowledge of human skeleton topology in the self-attention mechanism. To tackle this issue, in this paper, we propose a Pose-Oriented Transformer (POT) with uncertainty guided refinement for 3D HPE. Specifically, we first develop novel pose-oriented self-attention mechanism and distance-related position embedding for POT to explicitly exploit the human skeleton topology. The pose-oriented self-attention mechanism explicitly models the topological interactions between body joints, whereas the distance-related position embedding encodes the distance of joints to the root joint to distinguish groups of joints with different difficulties in regression. Furthermore, we present an Uncertainty-Guided Refinement Network (UGRN) to refine pose predictions from POT, especially for the difficult joints, by considering the estimated uncertainty of each joint with uncertainty-guided sampling strategy and self-attention mechanism. Extensive experiments demonstrate that our method significantly outperforms the state-of-the-art methods with reduced model parameters on 3D HPE benchmarks such as Human3.6M and MPI-INF-3DHP.

Fine-Grained Position Helps Memorizing More, a Novel Music Compound Transformer Model with Feature Interaction Fusion

Due to the particularity of the simultaneous occurrence of multiple events in music sequences, compound Transformer is proposed to deal with the challenge of long sequences. However, there are two deficiencies in the compound Transformer. First, since the order of events is more important for music than natural language, the information provided by the original absolute position embedding is not precise enough. Second, there is an important correlation between the tokens in the compound word, which is ignored by the current compound Transformer. Therefore, in this work, we propose an improved compound Transformer model for music understanding. Specifically, we propose an attribute embedding fusion module and a novel position encoding scheme with absolute-relative consideration. In the attribute embedding fusion module, different attributes are fused through feature permutation by using a multi-head self-attention mechanism in order to capture rich interactions between attributes. In the novel position encoding scheme, we propose RoAR position encoding, which realizes rotational absolute position encoding, relative position encoding, and absolute-relative position interactive encoding, providing clear and rich orders for musical events. Empirical study on four typical music understanding tasks shows that our attribute fusion approach and RoAR position encoding brings large performance gains. In addition, we further investigate the impact of masked language modeling and casual language modeling pre-training on music understanding.

PLSNet: Position-aware GCN-based autism spectrum disorder diagnosis via FC learning and ROIs sifting

Brain function connectivity, derived from functional magnetic resonance imaging (fMRI), has enjoyed high popularity in the studies of Autism Spectrum Disorder (ASD) diagnosis. Albeit rapid progress has been made, most studies still suffer from several knotty issues: (1) the hardship of modeling the sophisticated brain neuronal connectivity; (2) the mismatch of identically graph node setup to the variations of different brain regions; (3) the dimensionality explosion resulted from excessive voxels in each fMRI sample; (4) the poor interpretability giving rise to unpersuasive diagnosis. To ameliorate these issues, we propose a position-aware graph-convolution-network-based model, namely PLSNet, with superior accuracy and compelling built-in interpretability for ASD diagnosis. Specifically, a time-series encoder is designed for context-rich feature extraction, followed by a function connectivity generator to model the correlation with long range dependencies. In addition, to discriminate the brain nodes with different locations, the position embedding technique is adopted, giving a unique identity to each graph region. We then embed a rarefying method to sift the salient nodes during message diffusion, which would also benefit the reduction of the dimensionality complexity. Extensive experiments conducted on Autism Brain Imaging Data Exchange demonstrate that our PLSNet achieves state-of-the-art performance. Notably, on CC200 atlas, PLSNet reaches an accuracy of 76.4% and a specificity of 78.6%, overwhelming the previous state-of-the-art with 2.5% and 6.5% under five-fold cross-validation policy. Moreover, the most salient brain regions predicted by PLSNet are closely consistent with the theoretical knowledge in the medical domain, providing potential biomarkers for ASD clinical diagnosis. Our code is available at https://github.com/CodeGoat24/PLSNet.

HGR-ViT: Hand Gesture Recognition with Vision Transformer

Hand gesture recognition (HGR) is a crucial area of research that enhances communication by overcoming language barriers and facilitating human-computer interaction. Although previous works in HGR have employed deep neural networks, they fail to encode the orientation and position of the hand in the image. To address this issue, this paper proposes HGR-ViT, a Vision Transformer (ViT) model with an attention mechanism for hand gesture recognition. Given a hand gesture image, it is first split into fixed size patches. Positional embedding is added to these embeddings to form learnable vectors that capture the positional information of the hand patches. The resulting sequence of vectors are then served as the input to a standard Transformer encoder to obtain the hand gesture representation. A multilayer perceptron head is added to the output of the encoder to classify the hand gesture to the correct class. The proposed HGR-ViT obtains an accuracy of 99.98%, 99.36% and 99.85% for the American Sign Language (ASL) dataset, ASL with Digits dataset, and National University of Singapore (NUS) hand gesture dataset, respectively.

Hierarchical Wi-Fi Trajectory Embedding for Indoor User Mobility Pattern Analysis

The recent advances in smart building technologies have enabled us to collect massive Wi-Fi network based trajectory data, which provide an unparalleled opportunity for understanding the indoor user mobility pattern and enabling a wide range of business applications. While some previous studies have explored the Wi-Fi positioning of users, there still lacks a systematic and effective solution for indoor user mobility pattern analysis based on Wi-Fi trajectory data. To this end, in this paper, we propose a unified framework for modeling Wi-Fi trajectory data, namely HWTE, which can empower various tasks of indoor user mobility pattern analysis, such as user classification, next location prediction and schedule estimation. Specifically, we first propose a session trajectory construction module to extract the spatio-temporal semantic information from the Wi-Fi trajectories of users. Then, we devise a pre-training module to learn the unified representation of Wi-Fi trajectories. In particular, a session position embedding technique and a position query task is introduced to enhance the representation ability of the whole trajectory. Moreover, we further propose a hierarchical Transformer-based fine-tuning module to support various application tasks with time and space efficiency. Finally, we validate our framework on a real-world dataset with all three kinds of downstream tasks.

MCPT: Mixed Convolutional Parallel Transformer for Polarimetric SAR Image Classification

Vision transformers (ViT) have the characteristics of massive training data and complex model, which cannot be directly applied to polarimetric synthetic aperture radar (PolSAR) image classification tasks. Therefore, a mixed convolutional parallel transformer (MCPT) model based on ViT is proposed for fast PolSAR image classification. First of all, a mixed depthwise convolution tokenization is introduced. It replaces the learnable linear projection in the original ViT to obtain patch embeddings. The process of tokenization can reduce computational and parameter complexity and extract features of different receptive fields as input to the encoder. Furthermore, combining the idea of shallow networks with lower latency and easier optimization, a parallel encoder is implemented by pairing the same modules and recombining to form parallel blocks, which can decrease the network depth and computing power requirement. In addition, the original class embedding and position embedding are removed during tokenization, and a global average pooling layer is added after the encoder for category feature extraction. Finally, the experimental results on AIRSAR Flevoland and RADARSAT-2 San Francisco datasets show that the proposed method achieves a significant improvement in training and prediction speed. Meanwhile, the overall accuracy achieved was 97.9% and 96.77%, respectively.

GIT: A Transformer-Based Deep Learning Model for Geoacoustic Inversion

Geoacoustic inversion is a challenging task in marine research due to the complex environment and acoustic propagation mechanisms. With the rapid development of deep learning, various designs of neural networks have been proposed to solve this issue with satisfactory results. As a data-driven method, deep learning networks aim to approximate the inverse function of acoustic propagation by extracting knowledge from multiple replicas, outperforming conventional inversion methods. However, existing deep learning networks, mainly incorporating stacked convolution and fully connected neural networks, are simple and may neglect some meaningful information. To extend the network backbone for geoacoustic inversion, this paper proposes a transformer-based geoacoustic inversion model with additional frequency and sensor 2-D positional embedding to perceive more information from the acoustic input. The simulation experimental results indicate that our proposed model achieves comparable inversion results with the existing inversion networks, demonstrating its effectiveness in marine research.

What should be encoded by position embedding for neural network language models?

AbstractWord order is one of the most important grammatical devices and the basis for language understanding. However, as one of the most popular NLP architectures, Transformer does not explicitly encode word order. A solution to this problem is to incorporate position information by means of position encoding/embedding (PE). Although a variety of methods of incorporating position information have been proposed, the NLP community is still in want of detailed statistical researches on position information in real-life language. In order to understand the influence of position information on the correlation between words in more detail, we investigated the factors that affect the frequency of words and word sequences in large corpora. Our results show that absolute position, relative position, being at one of the two ends of a sentence and sentence length all significantly affect the frequency of words and word sequences. Besides, we observed that the frequency distribution of word sequences over relative position carries valuable grammatical information. Our study suggests that in order to accurately capture word–word correlations, it is not enough to focus merely on absolute and relative position. Transformers should have access to more types of position-related information which may require improvements to the current architecture.

Feature Correlation Transformer for Estimating Ambiguous Optical Flow

Cost volume is widely used to establish correspondences in optical flow estimation. However, when dealing with low-texture and occluded areas, it is difficult to estimate the cost volume correctly. Therefore, we propose a replacement: feature correlation transformer (FCTR), a transformer with self- and cross-attention alternations for obtaining global receptive fields and positional embedding for establishing correspondences. With global context and positional information, FCTR can produce more accurate correspondences for ambiguous areas. Using position-embedded feature allows the removal of the context network; the positional information can be aggregated within ambiguous motion boundaries, and the number of model parameters can be reduced. To speed up network convergence and strengthen robustness, we introduce a smooth L1 loss with exponential weights in the pre-training step. At the time of submission, our method achieves competitive performance with all published optical flow methods on both the KITTI-2015 and MPI-Sintel benchmarks. Moreover, it outperforms all optical flow and scene flow methods in KITTI-2015 foreground-region prediction.

Improved robustness of vision transformers via prelayernorm in patch embedding

Vision Transformers (ViTs) have recently demonstrated state-of-the-art performance in various vision tasks, replacing convolutional neural networks (CNNs). However, because ViT has a different architectural design than CNN, it may behave differently. To investigate whether ViT has a different performance or robustness, we tested ViT and CNN under various imaging conditions in practical vision tasks. We confirmed that for most image transformations, ViT’s robustness was comparable or even better than that of CNN. However, for contrast enhancement, ViT performed particularly poorly. We show that this is because positional embedding in ViT’s patch embedding can work improperly when the color scale changes. We demonstrate that the use of PreLayerNorm, a modified patch embedding structure, ensures the consistent behavior of ViT. Results demonstrate that ViT with PreLayerNorm exhibited improved robustness in the contrast-varying environments.

Improving Arabic Named Entity Recognition with a Modified Transformer Encoder

This article investigates the use of a transformer encoder for Arabic Named Entity Recognition (NER). The classic transformer that was originally proposed for machine translation adopts the absolute sinusoidal position embedding which is aware of distance but unfortunately is not aware of the directionality. However, in the NER task, both distance and orientation are crucial. Therefore, in this study, instead of using absolute sinusoidal position encoding, we employ relative positional encoding and incorporate the directionality information in our NER model. More specifically, our proposed model uses Bidirectional Long Short-Term Memory (BiLSTM) for encoding every input token. Then, the output of the encoder is fed to the multi-head attention where both the distance and directionality information are incorporated. The decoder layer with a simple fully connected layer takes as input, the result of the attention layer, and the prediction layer with Conditional Random Fields (CRF) predicts the tag of each token. We validate our proposed approach on two merged public datasets, namely, ANER corp and AQMAR. Our experiment results demonstrate significant improvements when compare to the vanilla Transformer with absolute sinusoidal position encoding while achieving a state-of-the-art result on a merged two Arabic public datasets.

SAT3D: Slot Attention Transformer for 3D Point Cloud Semantic Segmentation

Semantic segmentation of 3D point cloud is a key task in numerous intelligent transportation system applications, e.g. self-driving vehicles, traffic monitoring. Due to the sparsity and varying density of points in the outdoor point clouds, it becomes particularly challenging to extract object-centric features from data. This leads to poor semantic segmentation, especially for the rare object classes. To address that, we introduce the first-ever Slot Attention Transformer based technique to effectively model object-centric features in point cloud data. Our method uses cylindrical splits of space for voxelization and computes channel-wise positional embeddings before repetitively encoding the point cloud with slot attentions. Our second major contribution is a Large-Scale Outdoor Point Cloud dataset (SWAN), collected in a dense urban environment, driving 150km distance. It provides 16 billion points in more than 200K frames. The dataset also provides annotations for 10K frames for 24 classes. We also contribute a data augmentation scheme to handle rare object classes in real-world point clouds. Besides benchmarking popular existing methods on SWAN for the first time, we thoroughly evaluate our technique on the existing large-scale datasets, Semantic KITTI and nuScenes. Our results demonstrate a consistent performance gain for our technique, and verify the need of the more challenging SWAN dataset.

OPENnet: Object Position Embedding Network for Locating Anti-Bird Thorn of High-Speed Railway

OPENnet: Object Position Embedding Network for Locating Anti-Bird Thorn of High-Speed Railway