AIOZ AI @ CVPR 2022
✨ Coarse-to-Fine Reasoning for Visual Question Answering
🚀 Github: https://github.com/aioz-ai/CFR_VQA
Bridging the semantic gap between image and inquiry improves VQA accuracy. Most VQA approaches concentrate on attention mechanisms or visual relations to reason the response, whereas semantic characteristics are underutilized. In this paper, we provide a new reasoning framework to bridge visual and semantic VQA evidence. Our technique extracts image and question features and predicates. We present a new reasoning framework to learn coarse-to-fine features and predicates. Our methodology provides superior accuracy on three large-scale VQA datasets, compared to existing state-of-the-art methods. Our reasoning approach explains the deep neural network's answer prediction judgment.
AIOZ AI @ IV 2022 - The 33rd IEEE Intelligent Vehicles Symposium
✨ Deep Federated Learning for Autonomous Driving
🚀 Github: https://github.com/aioz-ai/FADNet
Autonomous driving is an active research topic in both academia and industry. Most present systems improve accuracy by training learnable models with centralized large-scale data. These approaches disregard user privacy. We provide a new way to understand autonomous driving policies while protecting privacy. We propose peer-to-peer Deep Federated Learning (DFL) to train deep architectures dispersed without central orchestration. We design a Federated Autonomous Driving network (FADNet) to improve model stability, ensure convergence, and tackle imbalanced data distribution concerns during federated learning. Our approach with FADNet and DFL yields greater accuracy compared to other previous methods. Our technique protects user privacy by not sending data to a central server.
AIOZ AI @ TMI - IEEE Transactions on Medical Imaging
✨ Light-weight deformable registration using adversarial learning with distilling knowledge
🚀 Github: https://github.com/aioz-ai/LDR_ALDK
Deformable registration is used in image-guided surgery and radiation therapy. Recent learning-based algorithms optimize non-linear spatial relationship between input images to improve accuracy. For real-time deployment, these technologies require current graphic cards. We introduce a Light-weight Deformable Registration network that minimizes computational cost while maintaining accuracy. We present a new adversarial learning algorithm that leverages important information from the effective but expensive instructor network to the student network. We make the student network lightweight and CPU-friendly. Extensive experiments on public datasets show that our suggested solution is accurate and faster than recent methods. Our adversarial learning system is vital for time-efficient deformable registration.