TAPe+ML v3

TAPe+ML v3 is a compact computer vision architecture for image classification, object detection, and instance segmentation. It uses TAPe (Theory of Active Perception) representations rather than raw pixels or fixed image patches. TAPe represents visual data as structured elements with known relationships. TAPe+ML v3 operates on these elements and their relationships through a set of small specialized submodels coordinated within one architecture. The system includes background separation, object localization, class prototype formation, and contour-based segmentation. The paper reports 88.1% Top-1 accuracy on ImageNet-1k and 89.9% on ImageNet-ReaL with fewer than 100K parameters. On COCO object detection, TAPe+ML v3 reports 84.7 mAP50 and 65.3 mAP50-95 with fewer than 100K parameters. On COCO instance segmentation, it reports 80.7 Mask mAP50 and 58.4 Mask mAP50-95. The paper also reports results on distribution-shift benchmarks: 78.6% on ObjectNet, 80.2% on ImageNet-v2, 90.3% on ImageNet-R, and 75.8% on ImageNet-Sketch. On Imagenette, the same 3-layer CNN architecture trained under identical conditions reached 92% accuracy on TAPe data versus 47% on raw pixels. In a DINO/iBOT experiment, the standard architecture reached a loss of 0.406 on TAPe data with 9K examples, while the raw-pixel baseline did not converge with 120K examples. TAPe was also evaluated for video scene detection. In the reported experiment, one hour of video was indexed in 10–11 seconds, the TAPe index was under 1 MB, and clustering took about one second. The paper includes an industrial domain-shift pilot. With 30 images, YOLO26s achieved approximately 28% accuracy, TAPe+ML in head-only mode achieved approximately 45%, and TAPe+ML with backbone adaptation achieved approximately 65%. With 85 images, backbone adaptation reached approximately 85% accuracy. With 500 images and NMS, the reported range was 85.7–96%. TAPe+ML also includes an auto-annotation workflow. After 10–15 images are labeled, the model can propose annotations for the remaining dataset; experts review and correct the proposed labels instead of annotating every object from scratch. Read the arXiv preprint: https://arxiv.org/abs/2609.20869 Project page: https://ml.comexp.net Contact: [email protected]