The TAPe+ML v3 model, detailed in a new arXiv paper, integrates classification, object detection, and instance segmentation into a single architecture with under 100,000 parameters. It achieves 88.1% Top-1 accuracy on ImageNet-1k and 65.3 mAP50–95 on COCO detection, outperforming larger SOTA models like YOLO and RF-DETR in efficiency.
In video indexing tasks, TAPe+ML processes one hour of footage in 10–11 seconds with an index size under 1 MB. This significantly outperforms alternatives like DINOv2, which takes approximately 5,220 seconds for indexing, and HOG, which generates an index of about 1.7 GB.
Designed for edge deployment, the full pipeline—including detection, classification, and segmentation—runs in approximately 12 ms per image on local hardware. The executable package size is around 10 MB, with peak memory usage during frame processing at about 20 MB.
The model demonstrates high performance in industrial OOD scenarios, such as detecting ore contamination on conveyor belts. On a 30-image pilot, TAPe+ML achieved 65% accuracy compared to 28% for YOLO26s, with performance scaling up to 96% as the dataset expanded to 500 images.
TAPe+ML includes an auto-annotation mode to reduce manual labeling requirements. This iterative process allows users to train models on small initial datasets and refine them by correcting the model's predictions, which is particularly useful for industrial applications requiring expert domain knowledge.
