Gain from Neighbors: Boosting Model Robustness in the Wild via Adversarial Perturbations Toward Neighboring Classes

Zhou Yang, Mingtao Feng, Tao Huang, Fangfang Wu, Weisheng Dong*, Xin Li, Guangming Shi

Project

Abstract: In this paper, we observe that classification errors arising from distribution shifts tend to cluster near the true values, suggesting that misclassifications commonly occur in semantically similar, neighboring categories. Furthermore, robust advanced vision foundation models maintain larger inter-class distances while preserving semantic consistency, making them less vulnerable to such shifts. Building on these findings, we propose a new method called GFN (Gain From Neighbors), which uses gradient priors from neighboring classes to perturb input images and incorporates an inter-class distance-weighted loss to improve class separation.

Vector Quantization with Self-Attention for Quality-Independent Representation Learning

Zhou Yang, Weisheng Dong*, Xin Li, Mengluan Huang, Yulin Sun and Guangming Shi

Project

Abstract: Inspired by sparse representation in image restoration, we opt to address the degraded image recognition problem by learning image-quality-independent feature representation in a simple plug-and-play manner, that is, to introduce discrete vector quantization (VQ) to remove redundancy in recognition models.

Self-feature Distillation with Uncertainty Modeling for Degraded Image Recognition

Zhou Yang, Weisheng Dong*, Xin Li, Jinjian Wu, Leida Li and Guangming Shi

Project

Abstract: In standard feature distillation, Mean Squared Error, MSE treats each pixel in the feature equally and may result in relatively poor reconstruction performance in some difficult regions. To address this issue, we propose a novel self-feature distillation method with uncertainty modeling for better producing HQ-like features from low-quality observations in this paper.

Inverse weight-balancing for deep long-tailed learning

Wenqi Dang, Zhou Yang, Weisheng Dong, Xin Li, Guangming Shi

Project

Abstract: We propose an inverse weight-balancing (IWB) approach to address data imbalance in deep learning. In the first stage, the encoder and classifier are trained with cross-entropy loss. In the second stage, the classifier is finetuned with an adaptive distribution for IWB. Unlike inverse image frequency methods, IWB aligns both class-wise and sample-wise distributions, improving performance on imbalanced datasets like CIFAR100-LT, ImageNet-LT, and iNaturalist2018.

BrickPal: Assembly as Language, Immersive and Gamification, In-situ Creation

Xiaofeng Zhang, Zhou Yang, Xiao Tang, Yao Shi, Hongni Ye, Yi Wu and Ran Zhang

Video

Introduction: We have developed an application using AR glasses for assisting assembly on the Unity Vuforia platform. This program also utilizes NLP technology to predict the next assembly step and supports user-defined assembly requirements. My main contribution is the assembly guidance using AR technology.

A Gesture-Based Human-Computer Interaction Shooting Game

Zhou Yang

Video

Introduction: Through combining a depth camera Kinect with object detection and gesture recognition algorithms, we can control the movement and shooting of a game characters based on the position and gestures of the hands in 3D space. Completed independently.

A 6-DoF Robotic Arm for Exploring Inverse Kinematics Solutions and Vision-Based Grasping.

Zhou Yang

Video

Introduction: A hardware platform for initial attempts at inverse kinematics solutions was built using six PWM servos and a 16-channel serial bus control board.