우수논문상(Best Paper Award) 수상 - 김정훈 석사과정

ICCSA 2026
The 26th International Conference on Computational Science and Its Applications
University of Minho, Braga, Portugal (June 30 - July 3, 2026)

"DeOP-Light: Analyzing the Fine-tuning Paradox in CLIP Visual Encoder Compression for Open-Vocabulary Semantic Segmentation"

doi: https://doi.org/10.1007/978-3-032-30488-9_12

Abstract. Open-vocabulary semantic segmentation (OVSS) relies on contrastive language-image pre-training (CLIP) visual encoders to clas- sify unseen categories at inference time. Compressing these encoders for efficient deployment is practically important, yet the impact on the crit- ical seen/unseen class distinction has not been studied. We present a diagnostic empirical study rather than a new architectural proposal of CLIP visual encoder compression within the decoupled one-pass (DeOP), applying floating-point (FP)16 quantization and low-rank adaptation (LoRA) on COCO-Stuff and Pascal VOC. We uncover a consistent asym- metric degradation: fine-tuning (FT) the CLIP visual encoder recovers seen class performance but simultaneously degrades unseen class perfor- mance, emerging within the first 1,000 iterations. By comparing against an uncompressed FT baseline, we show that FT itself not compression is the primary driver of this degradation. FT-free compression best pre- serves unseen performance, and post-hoc weight restoration partially re- covers it without retraining. Our findings provide practical guidelines for model compression in resource-constrained deployment and highlight the need for seen/unseen-aware evaluation in OVSS.

우수논문상을 수상한 논문 외에 3편의 논문이 발표되었습니다. 논문발표는 이지은 박사연구원, 김소연 석사연구원이 진행하였습니다.

1. C-MTSS: Compact Multi-teacher Single-Student Framework for Real-Time Multimodal Industrial Anomaly Detection (2026-06-30)
doi: https://doi.org/10.1007/978-3-032-30488-9_21

2. DepthGate: Confidence-Gated Depth Verification for Multimodal Industrial Anomaly Detection (2026-06-30)
doi: https://doi.org/10.1007/978-3-032-30488-9_7
3. GPS: GlobalCLIP-PatchCore-SAM Based Zero-Shot Anomaly Detection and Localization in Smart Manufacturing (2026-06-30)
doi: https://doi.org/10.1007/978-3-032-30488-9_28