Guo Haoren (郭昊仁)

Guo Haoren (郭昊仁)

Ph.D@NUS

Publications

You can also find my articles on my Google Scholar profile.

Journal

SDSimPoint: Shallow–Deep Similarity Learning for Few-Shot Point Cloud Semantic Segmentation

Published in TNNLS, 2025

Authors: Jiahui Wang, Haiyue Zhu , Haoren Guo , Abdullah Al Mamun , Cheng Xiang , Clarence W de Silva , Tong Heng Lee

Three-dimensional point cloud semantic segmentation is a core task in computer vision. While few-shot methods address limited-data scenarios better than fully supervised approaches, they often struggle to capture class-specific features due to class-agnostic pretraining. To overcome this, we propose SDSimPoint, a shallow–deep similarity learning network that models both shallow (e.g., geometry, color) and deep (e.g., context, semantics) similarities between support and query samples. We further introduce BEAM (Beyond-Episode Attention Module), which expands attention beyond a single episode using memory units, enhancing similarity modeling. Additionally, our learnable distance metric adapts to complex data distributions. SDSimPoint consistently outperforms baselines across multiple datasets in few-shot point cloud segmentation. Read more

Download Paper

Remaining Useful Life Prediction via Frequency Emphasizing Mix-Up and Masked Reconstruction

Published in IEEE Transaction on Artificial Intelligence (TAI), 2024

Authors: Haoren Guo, Haiyue Zhu , Jiahui Wang , Vadakkepat Prahlad , Weng Khuen Ho , Clarence W de Silva , Tong Heng Lee

This paper addresses the challenge of predicting Remaining Useful Lifetime (RUL) in machinery within Industry 4.0 by proposing a framework that improves data utilization through frequency-domain analysis and semi-supervised learning. The Frequency Emphasizing Mix-Up Module (FEMM) enhances feature extraction, while the Masked Autoencoder Reconstruction Auxiliary Learning (MARAL) utilizes unlabeled data from unrestricted domains. Evaluations on the C-MAPSS dataset show the model outperforms others, achieving significant accuracy improvements, particularly with minimal labeled data. The results demonstrate a notable reduction in RMSE and overall prediction error. Read more

Download Paper

Conference

Enhancing Multivariate Time-Series Domain Adaptation via Contrastive Frequency Graph Discovery and Language-Guided Adversary Alignment

Published in The Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25), 2024

Authors: Haoren Guo, Haiyue Zhu , Jiahui Wang , Vadakkepat Prahlad , Weng Khuen Ho , Tong Heng Lee

Unsupervised domain adaptation (UDA) minimizes reliance on labeled data by aligning features between a labeled source domain and an unlabeled target domain, making it effective for multivariate time series (MTS) prediction. Existing MTS UDA methods often overlook inter-series dependencies and frequency-domain insights. To address this, the proposed Frequency Graph Discovery (FGD) module uncovers shared frequency information, while the Frequency-Contextual Contrastive Learning (FCCL) framework aligns frequency-contextual representations for label-invariant prediction. Additionally, the Language-guided Adversary Alignment (LAA) module leverages large language models (LLMs) to enhance alignment using text-encoded labeled embeddings. This approach achieves state-of-the-art results on public MTS datasets. Read more

Download Paper

Few-Shot Point Cloud Semantic Segmentation for CAM/CAD via Feature Enhancement and Efficient Dual Attention

Published in IECON, 2023

Authors: Jiahui Wang, Haiyue Zhu , Haoren Guo , Abdullah Al Mamun , Clarence W De Silva , Tong Heng Lee

This paper presents a few-shot learning approach for 3D semantic segmentation in CAM/CAD workflows, aimed at improving the segmentation of novel classes in scenarios with high mixture but low volume. The method introduces Sequential Dual Attention (SDA) to capture both channel and spatial features, along with a non-parametric feature enhancement block to improve class recognition in the feature space. Compared to other few-shot models, the proposed approach achieves superior performance across multiple few-shot segmentation scenarios on two datasets, demonstrating significant improvements over baseline methods. Read more

Download Paper

Lightweight Compressed Temporal and Compressed Spatial Attention with Augmentation Fusion in Remaining Useful Life Prediction

Published in IECON, 2023

Authors: Haoren Guo, Haiyue Zhu , Jiahui Wang , Vadakkepat Prahlad , Weng Khuen Ho , Clarence W de Silva , Tong Heng Lee

This paper proposes a non-transformer model, Compressed Temporal and Compressed Spatial (CTCS) Attention, for predicting Remaining Useful Lifetime (RUL), addressing concerns with transformer-based models such as high computational complexity and ineffective handling of low data. The CTCS model efficiently captures temporal and spatial information with pre- and post-positional encodings, while an Augmentation Fusion Module (AFM) improves understanding of data invariances. Evaluated on the C-MAPSS dataset, the proposed model outperforms transformer-based methods in accuracy while reducing computational cost by up to 32 times fewer FLOPs. Read more

Download Paper

Few-shot point cloud semantic segmentation via contrastive self-supervision and multi-resolution attention

Published in ICRA, 2023

Authors: Jiahui Wang, Haiyue Zhu , Haoren Guo , Abdullah Al Mamun , Clarence W De Silva , Tong Heng Lee

This paper introduces a few-shot point cloud semantic segmentation approach optimized for real-world applications. Unlike existing methods that rely on fully-supervised pretraining with large annotated datasets, which leads to biased feature extraction, the proposed method employs a contrastive self-supervision framework for pretraining. This approach uses class-agnostic contrastive supervision and a learnable augmentor to eliminate feature bias, while a multi-resolution attention module and center-concentrated multi-prototype enhance feature extraction and mitigate intra-class sparsity. Experiments demonstrate state-of-the-art performance, with a case study showcasing its effectiveness in practical CAM/CAD segmentation. Read more

Download Paper

Masked Self-Supervision for Remaining Useful Lifetime Prediction in Machine Tools

Published in INDIN, 2022

Authors: Haoren Guo, Haiyue Zhu , Jiahui Wang , Prahlad Vadakkepat , Weng Khuen Ho , Tong Heng Lee

Predicting the Remaining Useful Lifetime (RUL) of machines is crucial for preventing tool wear and breakdown in Industry 4.0, but fully-supervised models struggle due to limited labeled data from broken machines. To address this, the authors propose a masked self-supervised learning method that uses unlabeled data for RUL prediction, showing superior performance on the C-MAPSS dataset compared to fully-supervised approaches. Read more

Download Paper

CAM/CAD Point Cloud Part Segmentation via Few-Shot Learning

Published in INDIN, 2022

Authors: Jiahui Wang, Haiyue Zhu , Haoren Guo , Abdullah Al Mamun , Prahlad Vadakkepat , Tong Heng Lee

3D part segmentation plays a critical role in improving the efficiency and reducing defects in CAM/CAD manufacturing workflows, particularly for CNC machinery. Traditional segmentation methods rely on fully-supervised learning, which requires large annotated datasets and struggles to generalize to new segmentation tasks. To address these limitations, the authors propose a few-shot learning approach that improves generalization and flexibility by using fewer samples. Their method, inspired by the attMPTI network, incorporates a multi-prototype approach with self-attention, along with transform net and center loss block, to better capture 3D features while reducing storage space and enhancing processing efficiency. Read more

Download Paper