Object Classification Tasks Research Articles

Prior probability biases perceptual choices by modulating the accumulation rate, rather than the baseline, of decision evidence

Abstract The prior probability of an upcoming stimulus has been shown to influence the formation of perceptual decisions. Computationally these effects have typically been attributed to changes in the starting point (i.e. baseline) of evidence accumulation in sequential sampling models. More recently, it has also been proposed that prior probability might additionally lead to changes in the rate of evidence accumulation. Here, we introduce a neurally-informed behavioural modelling approach to understand whether prior probability influences the starting point, the rate of evidence accumulation or both. To this end, we employ a well-established visual object categorization task for which two neural components underpinning participants’ choices have been characterised using single-trial analysis of the electroencephalogram. These components are reliable measures of trial-by-trial variability in the quality of the relevant decision evidence, which we use to constrain the estimation of a hierarchical drift diffusion model of perceptual choice. We find that, unlike previous computational accounts, constraining the model with the endogenous variability in the relevant decision evidence, results in prior probability effects being explained primarily by changes in the rate of evidence accumulation rather than changes in the starting point or a combination of both. Ultimately, our neurally-informed modelling approach helps disambiguate the mechanistic effect of prior probability on perceptual decision formation, suggesting that prior probability biases primarily the interpretation of sensory evidence towards the most likely stimulus.

Enhancing Underwater Object Detection and Classification Using Advanced Imaging Techniques: A Novel Approach with Diffusion Models

Underwater object detection and classification pose significant challenges due to environmental factors such as water turbidity and variable lighting conditions. This research proposes a novel approach that integrates advanced imaging techniques with diffusion models to address these challenges effectively, aligning with Sustainable Development Goal (SDG) 14: Life Below Water. The methodology leverages the Convolutional Block Attention Module (CBAM), Modified Swin Transformer Block (MSTB), and Diffusion model to enhance the quality of underwater images, thereby improving the accuracy of object detection and classification tasks. This study utilizes the TrashCan dataset, comprising diverse underwater scenes and objects, to validate the proposed method’s efficacy. This study proposes an advanced imaging technique YOLO (you only look once) network (AIT-YOLOv7) for detecting objects in underwater images. This network uses a modified U-Net, which focuses on informative features using a convolutional block channel and spatial attentions for color correction and a modified swin transformer block for resolution enhancement. A novel diffusion model proposed using modified U-Net with ResNet understands the intricate structures in images with underwater objects, which enhances detection capabilities under challenging visual conditions. Thus, AIT-YOLOv7 net precisely detects and classifies different classes of objects present in this dataset. These improvements are crucial for applications in marine ecology research, underwater archeology, and environmental monitoring, where precise identification of marine debris, biological organisms, and submerged artifacts is essential. The proposed framework advances underwater imaging technology and supports the sustainable management of marine resources and conservation efforts. The experimental results demonstrate that state-of-the-art object detection methods, namely SSD, YOLOv3, YOLOv4, and YOLOTrashCan, achieve mean accuracies (mAP@0.5) of 57.19%, 58.12%, 59.78%, and 65.01%, respectively, whereas the proposed AIT-YOLOv7 net reaches a mean accuracy (mAP@0.5) of 81.4% on the TrashCan dataset, showing a 16.39% improvement. Due to this improvement in the accuracy and efficiency of underwater object detection, this research contributes to broader marine science and technology efforts, promoting the better understanding and management of aquatic ecosystems and helping to prevent and reduce the marine pollution, as emphasized in SDG 14.

OPTICAL DEEP LEARNING LANDMINE DETECTION BASED ON LIMITED DATASET OF AERIAL IMAGERY

Landmine detection is one of the most innovative applications of unmanned aerial vehicles that became possible due to rapid development of both aerial vehicles equipped by different optical cameras and sensors using different physical principles, and object classification and detection methods, including machine learning and especially deep learning. Optical detection is an essential part of the overall landmine detection process that can be performed either separately or in combination with data processing from other types of cameras or sensors. The development of deep convolutional neural networks has dramatically changed the landscape of optical detection by making them de-facto choice number one for the majority of object classification, detection and segmentation tasks. However, the deterrent factor in the case of landmine detection is limited availability of appropriate data for training that different researchers try to overcome in different ways. The assessment of necessary amount of training data for any particular object detection problem still remains an experimental task. Despite several years of development in this area, still there is a shortage of research based on real landmine imagery obtained from unmanned aerial vehicles, so currently any public effort in this direction is valuable and works as an inspiration for new researchers. This paper describes such a study, namely its first iteration in which popular open-source tools are used to build detection pipeline and estimation of their efficiency is done using limited amount of data. It is shown that the problem of limited amount of training data can be effectively overcome by data augmentation and iterational process of training optical landmine detector is demonstrated. The effectiveness of open-source tools and libraries for neural networks training, object detection and dataset preparation is also demonstrated.

Screening COVID-19 from chest X-ray images by an optical diffractive neural network with the optimized F number

The COVID-19 pandemic continues to significantly impact people’s lives worldwide, emphasizing the critical need for effective detection methods. Many existing deep learning-based approaches for COVID-19 detection offer high accuracy but demand substantial computing resources, time, and energy. In this study, we introduce an optical diffractive neural network (ODNN-COVID), which is characterized by low power consumption, efficient parallelization, and fast computing speed for COVID-19 detection. In addition, we explore how the physical parameters of ODNN-COVID affect its diagnostic performance. We identify the F number as a key parameter for evaluating the overall detection capabilities. Through an assessment of the connectivity of the diffractive network, we established an optimized range of F number, offering guidance for constructing optical diffractive neural networks. In the numerical simulations, a three-layer system achieves an impressive overall accuracy of 92.64% and 88.89% in binary- and three-classification diagnostic tasks. For a single-layer system, the simulation accuracy of 84.17% and the experimental accuracy of 80.83% can be obtained with the same configuration for the binary-classification task, and the simulation accuracy is 80.19% and the experimental accuracy is 74.44% for the three-classification task. Both simulations and experiments validate that the proposed optical diffractive neural network serves as a passive optical processor for effective COVID-19 diagnosis, featuring low power consumption, high parallelization, and fast computing capabilities. Furthermore, ODNN-COVID exhibits versatility, making it adaptable to various image analysis and object classification tasks related to medical fields owing to its general architecture.

Human Visual Cortex and Deep Convolutional Neural Network Care Deeply about Object Background.

Deep convolutional neural networks (DCNNs) are able to partially predict brain activity during object categorization tasks, but factors contributing to this predictive power are not fully understood. Our study aimed to investigate the factors contributing to the predictive power of DCNNs in object categorization tasks. We compared the activity of four DCNN architectures with EEG recordings obtained from 62 human participants during an object categorization task. Previous physiological studies on object categorization have highlighted the importance of figure-ground segregation-the ability to distinguish objects from their backgrounds. Therefore, we investigated whether figure-ground segregation could explain the predictive power of DCNNs. Using a stimulus set consisting of identical target objects embedded in different backgrounds, we examined the influence of object background versus object category within both EEG and DCNN activity. Crucially, the recombination of naturalistic objects and experimentally controlled backgrounds creates a challenging and naturalistic task, while retaining experimental control. Our results showed that early EEG activity (< 100 msec) and early DCNN layers represent object background rather than object category. We also found that the ability of DCNNs to predict EEG activity is primarily influenced by how both systems process object backgrounds, rather than object categories. We demonstrated the role of figure-ground segregation as a potential prerequisite for recognition of object features, by contrasting the activations of trained and untrained (i.e., random weights) DCNNs. These findings suggest that both human visual cortex and DCNNs prioritize the segregation of object backgrounds and target objects to perform object categorization. Altogether, our study provides new insights into the mechanisms underlying object categorization as we demonstrated that both human visual cortex and DCNNs care deeply about object background.

Deep Learning Based Object Detection with Unmanned Aerial Vehicle Equipped with Embedded System

This study aims to introduce an Unmanned Aerial Vehicle (UAV) platform capable of performing real-time object detection and classification tasks using computer vision techniques in the field of artificial intelligence. Previous scientific research reveals the utilization of two different methods for object detection and classification via UAVs. One of these methods involves transmitting the acquired UAV images to a ground control center for processing, whereafter the processed data is relayed back to the UAV. The other approach entails transferring images over the internet to a cloud system, where image processing is conducted, and the resultant data is subsequently sent back to the UAV. This allows the UAV to autonomously perform predefined tasks. Enabling the UAV with autonomous decision-making capabilities and the ability to perform object detection and classification from recorded images requires an embedded artificial intelligence module. The ability of the UAV to utilize image processing technologies through embedded systems significantly enhances its object detection and classification capabilities, providing it with a significant advantage. This enables the UAV to be used more effectively and reliably in various tasks. In the proposed approach, image processing was achieved by mounting a Raspberry Pi 4 and camera on the UAV. Additionally, a Raspberry Pi-compatible 4G/LTE modem kit was used to provide remote intervention capability, and the Coral Edge TPU auxiliary processor was used to increase object detection speed. The TensorFlow Library and the SSD MobilNetV2 convolutional neural network model were used for image processing. During test flights, accuracy values of approximately 96.3% for car detection and 96.2% for human detection were achieved.

Analysis of the PointNet neural network architecture

Objective. Most researchers convert point cloud data into ordinary three-dimensional voxel grids or image collections, which makes the data unnecessarily voluminous and causes problems when processing them. The purpose of the study is to analyze the architecture of the PointNet neural network. Method. A unified approach has been applied to solving various 3D recognition problems, ranging from object classification, detail segmentation to semantic scene analysis. Result. A comparative analysis of the classification of 2d and 3d objects was carried out, the layers and functions through which classification occurs were studied in detail. A type of neural network is considered that directly uses point clouds, which takes into account the invariance of permutations of points in the input data. The network is determined to provide a unified architecture for applications ranging from object classification, part segmentation, and scene semantics. For semantic segmentation, the input data can be either a single object from the part area segmentation or a small part of the 3D scene. A neural network that is widely used for raster image editing, graphic design, and digital art is a deep point cloud architecture called PointNet. Conclusion. A new deep point cloud architecture, PointNet, is introduced. For object classification task, the input point cloud is directly selected from the shape or pre-segmented from the scene point cloud. To obtain a virtual model of the real world, neural network solutions are used, based on the assumption that there is an RGB point cloud obtained by an RGB-D camera from one or several angles.

Dynamic convolution layer based optimization techniques for object classification and semantic segmentation

&lt;p&gt;Providing meaningful classification for each pixel in an image is a primary goal of computer vision, and the tasks of object classification and semantic segmentation are among the field’s greatest challenges. To improve object classification, this study presents a novel method that combines semantic segmentation with dynamic convolution layer-based optimization techniques. In the proposed method, a Refined Convolution Neural Network (R-CNN) is used, which uses non-extensive entropy to dynamically increase the size of its convolutional layers. The Common Objects in Context (COCO) dataset is used to assess the performance of the model. The model performs exceptionally well at different Intersections over Union (IoU) cutoffs, with average precision values of 40.1, 61.9, and 45.4, respectively, for Average Precision (AP), AP&lt;sub&gt;50&lt;/sub&gt;, and AP&lt;sub&gt;75&lt;/sub&gt;. These results demonstrate the model’s efficiency in discriminating between various image contents. Additionally, the model predicts an image’s outcome on average in just 0.901 s. The model has been proven to be superior through various performance evaluation parameters, showing an average mean precision of 91.78%. This study demonstrates the power of combining dynamic convolution layers with semantic segmentation to improve object classification accuracy, a key component in the development of computer vision applications.&lt;/p&gt;

TriCI: Triple Cross-Intra Branch Contrastive Learning for Point Cloud Analysis.

Whereas contrastive learning eliminates the need for labeled data, existing methods may suffer from inadequate features due to the conventional single shared encoder structure and struggle to fully harness the rich spectrum of 3D augmentations. In this paper, we propose TriCI, a self-supervised method that designs a triple-branch contrastive learning architecture. During contrastive pre-training, we generate three augmented versions of each input point cloud sample and pair each augmented sample with the original one, resulting in three unique positive pairs. We subsequently feed the pairs into three distinct encoders, each of which extracts features from its corresponding input positive pair. We design a novel cross-branch contrastive loss and use it along with the intra-branch contrastive loss to jointly train our network. The proposed cross-branch loss effectively aligns the output features from different perspectives for pre-training and facilitates their integration for downstream tasks, particularly in object-level scenarios. The intra-branch loss helps maximize the feature correspondences within positive pairs. Extensive experiments demonstrate the superiority of our TriCI in self-supervised learning, and show its strong ability in enhancing the performance of downstream object classification and part segmentation tasks. Interestingly, our TriCI achieves a 92.9% accuracy for linear SVM evaluation on ModelNet40, exceeding its closest competitor by 1.7% and even exceeding some supervised methods.

MLGCN: an ultra efficient graph convolutional neural model for 3D point cloud analysis.

With the rapid advancement of 3D acquisition technologies, 3D sensors such as LiDARs, 3D scanners, and RGB-D cameras have become increasingly accessible and cost-effective. These sensors generate 3D point cloud data that require efficient algorithms for tasks such as 3D model classification and segmentation. While deep learning techniques have proven effective in these areas, existing models often rely on complex architectures, leading to high computational costs that are impractical for real-time applications like augmented reality and robotics. In this work, we propose the Multi-level Graph Convolutional Neural Network (MLGCN), an ultra-efficient model for 3D point cloud analysis. The MLGCN model utilizes shallow Graph Neural Network (GNN) blocks to extract features at various spatial locality levels, leveraging precomputed KNN graphs shared across GCN blocks. This approach significantly reduces computational overhead and memory usage, making the model well-suited for deployment on low-memory and low-CPU devices. Despite its efficiency, MLGCN achieves competitive performance in object classification and part segmentation tasks, demonstrating results comparable to state-of-the-art models while requiring up to a thousand times fewer floating-point operations and significantly less storage. The contributions of this paper include the introduction of a lightweight, multi-branch graph-based network for 3D shape analysis, the demonstration of the model's efficiency in both computation and storage, and a thorough theoretical and experimental evaluation of the model's performance. We also conduct ablation studies to assess the impact of different branches within the model, providing valuable insights into the role of specific components.

Impact of aversive affect on neural mechanisms of categorization decisions.

Many theories contend that evidence accumulation is a critical component of decision-making. Cognitive accumulation models typically interpret two main parameters: a drift rate and decision threshold. The former is the rate of accumulation, based on the quality of evidence, and the latter is the amount of evidence required for a decision. Some studies have found neural signals that mimic evidence accumulators and can be described by the two parameters. However, few studies have related these neural parameters to experimental manipulations of sensory data or memory representations. Here, we investigated the influence of affective salience on neural accumulation parameters. High affective salience has been repeatedly shown to influence decision-making, yet its effect on neural evidence accumulation has been unexamined. The current study used a two-choice object categorization task of body images (feet or hands). Half the images in each category were high in affective salience because they contained highly aversive features (gore and mutilation). To study such quick categorization decisions with a relatively slow technique like functional magnetic resonance imaging, we used a gradual reveal paradigm to lengthen cognitive processing time through the gradual "unmasking" of stimuli. Because the aversive features were task-irrelevant, high affective salience produced a distractor effect, slowing decision time. In visual accumulation regions of interest, high affective salience produced a longer time to peak activation. Unexpectedly, the later peak appeared to be the product of changes to both drift rate and decision threshold. The drift rate for high affective salience was shallower, and the decision threshold was greater. To our knowledge, this is the first demonstration of an experimental manipulation of sensory data or memory representations that changed the neural decision threshold. These findings advance our knowledge of the neural mechanisms underlying affective responses in general and the influence of high affective salience on object representations and categorization decisions.

IntelPVT: intelligent patch-based pyramid vision transformers for object detection and classification

Since the advent of Transformers followed by Vision Transformers (ViTs), enormous success has been achieved by researchers in the field of computer vision and object detection. The difficulty mechanism of splitting images into fixed patches posed a serious challenge in this arena and resulted in loss of useful information at the time of object detection and classification. To overcome the challengers, we propose an innovative Intelligent-based patching mechanism and integrated it seamlessly into the conventional Patch-based ViT framework. The proposed method enables the utilization of patches with flexible sizes to capture and retain essential semantic content from input images and therefore increases the performance compared with conventional methods. Our method was evaluated with three renowned datasets Microsoft Common Objects in Context (MSCOCO-2017), Pascal VOC (Visual Object Classes Challenge) and Cityscapes upon object detection and classification. The experimental results showed promising improvements in specific metrics, particularly in higher confidence thresholds, making it a notable performer in object detection and classification tasks.

Self-supervised Multi-view Learning via Auto-encoding 3D Transformations

3D object representation learning is a fundamental challenge in computer vision to infer about the 3D world. Recent advances in deep learning have shown their efficiency in 3D object recognition, among which view-based methods have performed best so far. However, feature learning of multiple views in existing methods is mostly performed in a supervised fashion, which often requires a large amount of data labels with high costs. In contrast, self-supervised learning aims to learn multi-view feature representations without involving labeled data. To this end, we propose a novel self-supervised framework to learn Multi-View Transformation Equivariant Representations (MV-TER), exploring the equivariant transformations of a 3D object and its projected multiple views that we derive. Specifically, we perform a 3D transformation on a 3D object and obtain multiple views before and after the transformation via projection. Then, we train a representation encoding module to capture the intrinsic 3D object representation by decoding 3D transformation parameters from the fused feature representations of multiple views before and after the transformation. Experimental results demonstrate that the proposed MV-TER significantly outperforms the state-of-the-art view-based approaches in 3D object classification and retrieval tasks and show the generalization to real-world datasets. The code is available at https://github.com/gyshgx868/mvter .

Detection, instance segmentation, and classification for astronomical surveys with deep learning (deepdisc): detectron2 implementation and demonstration with Hyper Suprime-Cam data

ABSTRACT The next generation of wide-field deep astronomical surveys will deliver unprecedented amounts of images through the 2020s and beyond. As both the sensitivity and depth of observations increase, more blended sources will be detected. This reality can lead to measurement biases that contaminate key astronomical inferences. We implement new deep learning models available through Facebook AI Research’s detectron2 repository to perform the simultaneous tasks of object identification, deblending, and classification on large multiband co-adds from the Hyper Suprime-Cam (HSC). We use existing detection/deblending codes and classification methods to train a suite of deep neural networks, including state-of-the-art transformers. Once trained, we find that transformers outperform traditional convolutional neural networks and are more robust to different contrast scalings. Transformers are able to detect and deblend objects closely matching the ground truth, achieving a median bounding box Intersection over Union of 0.99. Using high-quality class labels from the Hubble Space Telescope, we find that when classifying objects as either stars or galaxies, the best-performing networks can classify galaxies with near 100 per cent completeness and purity across the whole test sample and classify stars above 60 per cent completeness and 80 per cent purity out to HSC i-band magnitudes of 25 mag. This framework can be extended to other upcoming deep surveys such as the Legacy Survey of Space and Time and those with the Roman Space Telescope to enable fast source detection and measurement. Our code, deepdisc, is publicly available at https://github.com/grantmerz/deepdisc.

The tortoise and the hare: Fast and slow learners in an object categorization task

The tortoise and the hare: Fast and slow learners in an object categorization task

R$^{2}$Fed: Resilient Reinforcement Federated Learning for Industrial Applications

Federated learning has become an emerging hot research field in industry because of its ability to perform large-scale distributed learning while preserving data privacy. However, recent studies have shown that in the actual use of federated learning, there are device heterogeneity and data Non-IID (Not Identically and Independently Distributed) characteristics between client nodes, which will affect the effect of federated learning. In this work we propose R <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula> Fed (Resilient Reinforcement Federated Learning), a resilient reinforcement federated learning method, which applies reinforcement learning to federated learning and uses reinforcement learning for weighted fusion of client models instead of average fusion. We conduct experiments on object detection, object classification, and sentiment classification tasks in the context of Non-IID and heterogeneity, and the experimental results show that the R <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula> Fed method outperforms traditional federated learning, increasing the average accuracy by 4.7%. Experiments also demonstrate that R <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula> Fed is resilient to federation attacks.

High-Performance and Lightweight AI Model for Robot Vacuum Cleaners with Low Bitwidth Strong Non-Uniform Quantization

Artificial intelligence (AI) plays a critical role in the operation of robot vacuum cleaners, enabling them to intelligently navigate to clean and avoid indoor obstacles. Due to limited computational resources, manufacturers must balance performance and cost. This necessitates the development of lightweight AI models that can achieve high performance. Traditional uniform weight quantization assigns the same number of levels to all weights, regardless of their distribution or importance. Consequently, this lack of adaptability may lead to sub-optimal quantization results, as the quantization levels do not align with the statistical properties of the weights. To address this challenge, in this work, we propose a new technique called low bitwidth strong non-uniform quantization, which largely reduces the memory footprint of AI models while maintaining high accuracy. Our proposed non-uniform quantization method, as opposed to traditional uniform quantization, aims to align with the actual weight distribution of well-trained neural network models. The proposed quantization scheme builds upon the observation of weight distribution characteristics in AI models and aims to leverage this knowledge to enhance the efficiency of neural network implementations. Additionally, we adjust the input image size to reduce the computational and memory demands of AI models. The goal is to identify an appropriate image size and its corresponding AI models that can be used in resource-constrained robot vacuum cleaners while still achieving acceptable accuracy on the object classification task. Experimental results indicate that when compared to the state-of-the-art AI models in the literature, the proposed AI model achieves a 2-fold decrease in memory usage from 15.51 MB down to 7.68 MB while maintaining the same accuracy of around 93%. In addition, the proposed non-uniform quantization model reduces memory usage by 20 times (from 15.51 MB down to 0.78 MB) with a slight accuracy drop of 3.11% (the classification accuracy is still above 90%). Thus, our proposed high-performance and lightweight AI model strikes an excellent balance between model complexity, classification accuracy, and computational resources for robot vacuum cleaners.

Deep Learning for Detecting Verticillium Fungus in Olive Trees: Using YOLO in UAV Imagery

The verticillium fungus has become a widespread threat to olive fields around the world in recent years. The accurate and early detection of the disease at scale could support solving the problem. In this paper, we use the YOLO version 5 model to detect verticillium fungus in olive trees using aerial RGB imagery captured by unmanned aerial vehicles. The aim of our paper is to compare different architectures of the model and evaluate their performance on this task. The architectures are evaluated at two different input sizes each through the most widely used metrics for object detection and classification tasks (precision, recall, mAP@0.5 and mAP@0.5:0.95). Our results show that the YOLOv5 algorithm is able to deliver good results in detecting olive trees and predicting their status, with the different architectures having different strengths and weaknesses.

Implementation of Field-Programmable Gate Array Platform for Object Classification Tasks Using Spike-Based Backpropagated Deep Convolutional Spiking Neural Networks.

This paper investigates the performance of deep convolutional spiking neural networks (DCSNNs) trained using spike-based backpropagation techniques. Specifically, the study examined temporal spike sequence learning via backpropagation (TSSL-BP) and surrogate gradient descent via backpropagation (SGD-BP) as effective techniques for training DCSNNs on the field programmable gate array (FPGA) platform for object classification tasks. The primary objective of this experimental study was twofold: (i) to determine the most effective backpropagation technique, TSSL-BP or SGD-BP, for deeper spiking neural networks (SNNs) with convolution filters across various datasets; and (ii) to assess the feasibility of deploying DCSNNs trained using backpropagation techniques on low-power FPGA for inference, considering potential configuration adjustments and power requirements. The aforementioned objectives will assist in informing researchers and companies in this field regarding the limitations and unique perspectives of deploying DCSNNs on low-power FPGA devices. The study contributions have three main aspects: (i) the design of a low-power FPGA board featuring a deployable DCSNN chip suitable for object classification tasks; (ii) the inference of TSSL-BP and SGD-BP models with novel network architectures on the FPGA board for object classification tasks; and (iii) a comparative evaluation of the selected spike-based backpropagation techniques and the object classification performance of DCSNNs across multiple metrics using both public (MNIST, CIFAR10, KITTI) and private (INHA_ADAS, INHA_KLP) datasets.

PaRot: Patch-Wise Rotation-Invariant Network via Feature Disentanglement and Pose Restoration

Recent interest in point cloud analysis has led rapid progress in designing deep learning methods for 3D models. However, state-of-the-art models are not robust to rotations, which remains an unknown prior to real applications and harms the model performance. In this work, we introduce a novel Patch-wise Rotation-invariant network (PaRot), which achieves rotation invariance via feature disentanglement and produces consistent predictions for samples with arbitrary rotations. Specifically, we design a siamese training module which disentangles rotation invariance and equivariance from patches defined over different scales, e.g., the local geometry and global shape, via a pair of rotations. However, our disentangled invariant feature loses the intrinsic pose information of each patch. To solve this problem, we propose a rotation-invariant geometric relation to restore the relative pose with equivariant information for patches defined over different scales. Utilising the pose information, we propose a hierarchical module which implements intra-scale and inter-scale feature aggregation for 3D shape learning. Moreover, we introduce a pose-aware feature propagation process with the rotation-invariant relative pose information embedded. Experiments show that our disentanglement module extracts high-quality rotation-robust features and the proposed lightweight model achieves competitive results in rotated 3D object classification and part segmentation tasks.