TAPe+ML: A Compact Structured Representation for Multi-Task Computer Vision
TAPe+ML v3 reports COCO detection and segmentation results with fewer than 100,000 parameters.
The paper describes a vision system that shifts work from raw pixel tensors into a shared structured TAPe representation before recognition. Its modular setup covers image classification, object detection, and instance segmentation, with reported COCO scores of 84.7 mAP50 for detection and 80.7 mask mAP50 for segmentation. The authors also report 92% validation accuracy on Imagenette and 89.9% Top-1 on ImageNet-Real. They argue the representation can reduce data, memory, and compute needs for multi-task vision systems. HF Daily Papers' note
The paper describes a vision system that shifts work from raw pixel tensors into a shared structured TAPe representation before recognition. Its modular setup covers image classification, object detection, and instance segmentation, with reported COCO scores of 84.7 mAP50 for detection and 80.7 mask mAP50 for segmentation. The authors also report 92% validation accuracy on Imagenette and 89.9% Top-1 on ImageNet-Real. They argue the representation can reduce data, memory, and compute needs for multi-task vision systems. HF Daily Papers' note
score 5