1) 𝐔𝐬𝐢𝐧𝐠 𝐌𝐈𝐌𝐎 𝐦𝐦𝐖𝐚𝐯𝐞 𝐫𝐚𝐝𝐚𝐫 𝐟𝐨𝐫 𝐎𝐛𝐣𝐞𝐜𝐭 𝐃𝐞𝐭𝐞𝐜𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐭𝐫𝐚𝐜𝐤𝐢𝐧𝐠 𝐩𝐫𝐨𝐨𝐟-𝐨𝐟-𝐜𝐨𝐧𝐜𝐞𝐩𝐭• Developed a tool to collect raw data from TI MIMO mmWave radar and synchronize radar frames with depth camera frames• Used YOLOv8 and MOT for automated data annotation for object detection• Trained U-Net model for semantic segmentation of radar images• Implemented a custom CNN model for object detection with radar images2) 𝐒𝐞𝐪𝐮𝐞𝐧𝐜𝐞-𝐭𝐨-𝐬𝐞𝐪𝐮𝐞𝐧𝐜𝐞 𝐠𝐞𝐬𝐭𝐮𝐫𝐞𝐬 𝐜𝐥𝐚𝐬𝐬𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧• Proposed and implemented a two-stage training scheme that uses metric learning to improve model robustness - twice lower false positive rate and ~+5% accuracy3) 𝐇𝐮𝐦𝐚𝐧 𝐩𝐫𝐞𝐬𝐞𝐧𝐜𝐞 𝐝𝐞𝐭𝐞𝐜𝐭𝐢𝐨𝐧• Developed a SincNet-based model for human presence detection that can reliably classify between a moving background and a sitting human• Used Metric Learning to improve the model’s generalization and to filter outliers - achieved balanced accuracy over 96.6% on a hard dataset (moving background and sitting/standing still human)

Projects:1) 𝐂𝐫𝐨𝐬𝐬-𝐩𝐥𝐚𝐭𝐟𝐨𝐫𝐦 𝐌𝐨𝐛𝐢𝐥𝐞 𝐀𝐮𝐠𝐦𝐞𝐧𝐭𝐞𝐝 𝐑𝐞𝐚𝐥𝐢𝐭𝐲 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤 (Excellence Collection Co., Ltd.; Farsail Technology Co., Ltd.) • Developed a cross-platform AR framework for iOS and Android from scratch, focusing on planar and 3D object tracking • Achieved 30 FPS with a 720p camera resolution on mobile phones, ensuring smooth and responsive AR experiences2) 𝐓𝐫𝐚𝐟𝐟𝐢𝐜 𝐌𝐨𝐧𝐢𝐭𝐨𝐫𝐢𝐧𝐠 𝐚𝐧𝐝 𝐀𝐧𝐚𝐥𝐲𝐬𝐢𝐬 (Institute for Information Industry, Taiwan) • Implemented Multiple Objects Tracking (MOT) system with vehicle behavior classification and prediction for traffic cameras • Developed a neural network model for predicting vehicles and pedestrian trajectories, achieving a prediction accuracy of 0.29 meters for 3-second trajectories and outperforming other models on the inD dataset3) 𝐈𝐦𝐚𝐠𝐞 𝐒𝐭𝐢𝐭𝐜𝐡𝐢𝐧𝐠 (Laser Solution Technology Co., Ltd) • Implemented a custom image stitcher for an industrial camera • Achieved faster than 1 second processing time on CPU using multithreading (for up to 35 input images with 1280x1024 resolution) • Delivered pixel-accurate image stitching with image overlap < 30%4) 𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐢𝐜 𝐎𝐩𝐭𝐢𝐜𝐚𝐥 𝐈𝐧𝐬𝐩𝐞𝐜𝐭𝐢𝐨𝐧 (ABC Taiwan Electronics Corp, PCB inductors) (team of 3, senior software engineer) • Developed an active learning pipeline for fast dataset annotation • Enhanced defect detection accuracy by approximately 50% in YOLOv5 models through task-specific image preprocessing • Achieved defect detection accuracy surpassing that of trained specialists5) 𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐢𝐜 𝐎𝐩𝐭𝐢𝐜𝐚𝐥 𝐈𝐧𝐬𝐩𝐞𝐜𝐭𝐢𝐨𝐧 (Elytone Electronic Co., Ltd., PCB transformers) (team of 6, senior software engineer) • used instance segmentation CNN (SOLOv2) for PCB transformers segmentation and integrated it with classical image processing algorithms for defect detection • Achieved 100% defect detection accuracy on test samples