EdgeCrafter: Compact ViTs for Edge Dense Prediction via Task-Specialized Distillation

Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• EdgeCrafter introduces ECDet, a detection model built from a distilled compact backbone and an edge-friendly encoder-decoder design.method 95%
• ECDet-S achieves 51.7 AP on COCO with fewer than 10M parameters using only COCO annotations.result 95%
• ECInsSeg achieves instance segmentation performance comparable to RF-DETR while using substantially fewer parameters.result 90%
• ECPose-X reaches 74.8 AP for pose estimation, significantly outperforming YOLO26Pose-X (71.6 AP).result 95%
• Compact Vision Transformers struggle to achieve strong accuracy-efficiency tradeoffs compared to CNN-based architectures like YOLO for edge dense prediction.limitation 90%
• The performance gap between compact ViTs and CNNs for edge dense prediction is due to insufficient task-specific representation learning in small-scale ViTs.technical 85%
• Compact ViTs with task-specialized distillation and edge-aware design can be a practical and competitive option for edge dense prediction.result 90%
• YOLO26Pose-X relies on extensive Objects365 pretraining for its pose estimation performance.technical 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card