기본 정보
연구 분야
프로젝트
논문
구성원
article|
gold
·인용수 0
·2026
Edge-VIS: Knowledge-Distilled Video Instance Segmentation for Energy-Efficient Real-Time Robotic Scene Understanding
Hasnain Hyder, Yong-Woon Kim, ChulUng Kang, Yung-Cheol Byun
IEEE Access
초록

Video instance segmentation (VIS) is crucial for robotic perception but state-of-the-art transformer-based methods require computational resources for edge deployment. We propose Edge-VIS, the first knowledge distillation framework specifically designed for video instance segmentation, addressing the previously unexplored challenge of transferring sophisticated spatiotemporal reasoning, spatial segmentation fidelity, and temporal instance tracking consistency simultaneously from a transformer-based teacher (GenVIS-Swin-L, 1,229 M parameters) to a lightweight CNN-based student (modified YOLOv8n-seg, 2.7 M parameters) for real-time edge deployment in robotic systems. Our multilevel distillation strategy includes feature-level, reaction-level, and attention-based distillation to preserve both partition quality and temporal stability. On the challenging Occluded Video Instance Segmentation (OVIS) dataset emphasizing severe constraints, Edge-VIS achieves 46.0% mAP@0.5, while requiring only 14 GFLOPs (36.4% reduction) and 2.7 M parameters (15.6% reduction from baseline YOLOv8n-seg). With respect to the teacher model, Edge-VIS achieves 66.5% accuracy while being 455x more parameter efficient and achieving 10.4x faster inference (52 vs 5 FPS). The deployment on NVIDIA Jetson AGX Orin demonstrates real-time performance at 250 FPS with only 29 W power consumption, confirming practical feasibility for battery-powered robotic systems.

키워드
SegmentationImage segmentationScale-space segmentationRobotFeature extractionMobile robot
타입
article
IF / 인용수
- / 0
게재 연도
2026

주식회사 디써클

대표 장재우,이윤구서울특별시 강남구 역삼로 169, 명우빌딩 2층 (TIPS타운 S2)대표 전화 0507-1312-6417이메일 info@rndcircle.io사업자등록번호 458-87-03380호스팅제공자 구글 클라우드 플랫폼(GCP)

© 2026 RnDcircle. All Rights Reserved.