TAPe+ML v3: A Compact Structured Representation for Multi-Task Computer Vision

Authors: Sergey Kurinov, Alexey Upatov

Affiliation: Comexp Research Lab, TAPe + ML Project, Nizhniy Novgorod, Russia

arXiv: 2609.20869 [cs.CV] — Submitted 15 Sep 2026

Subjects: Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV)

DOI: https://doi.org/10.48550/arXiv.2609.20869 (pending DataCite registration)

Abstract

We present TAPe+ML v3, a compact computer vision system built on TAPe (Theory of Active Perception), a structured representation that encodes relations among perceptual elements before recognition. Rather than operating directly on pixel tensors, the system uses a shared TAPe representation and a modular recognition architecture that supports image classification, object detection, and instance segmentation.

TAPe+ML v3 combines background and contour processing, local object localization, prototype-based classification, and a coordinator for specialized submodels. Across the reported experiments, it uses fewer than 100,000 parameters.

On COCO object detection, it achieves 84.7 mAP50 and 65.3 mAP50-95. On COCO instance segmentation, it reaches 80.7 mask mAP50 and 58.4 mask mAP50-95. In classification experiments, it attains 92% validation accuracy on Imagenette under an identical-training comparison with a raw-pixel baseline, and 89.9% Top-1 accuracy on ImageNet-Real. The authors also evaluate compactness in video scene detection and adaptation under distribution shift in an industrial pilot.

These results suggest that shifting part of the modeling burden from network parameters to a structured input representation can support compact multi-task vision systems with reduced data, memory, and compute requirements.

Why This Matters in 2026

As foundation models continue to scale, TAPe+ML v3 moves in the opposite direction—toward structured, parameter-efficient multi-task vision. In 2026, deployment constraints (edge devices, industrial pilots, on-prem inference) are driving renewed interest in systems that trade raw parameter count for structured priors. TAPe+ML v3's sub-100K-parameter footprint, shared representation across three tasks, and demonstrated robustness under distribution shift make it a relevant data point for the "small but structured" branch of modern computer vision.

Key Contributions

Structured pre-recognition representation — TAPe encodes relations among perceptual elements prior to recognition, rather than learning directly from raw pixel tensors.
Shared representation, modular recognition — One TAPe representation feeds a modular architecture covering classification, detection, and instance segmentation.
Extreme parameter efficiency — Under 100,000 parameters across all reported tasks.
Strong benchmark results — 84.7 mAP50 / 65.3 mAP50-95 on COCO detection; 80.7 / 58.4 mask mAP on COCO instance segmentation; 92% Imagenette validation accuracy; 89.9% ImageNet-Real Top-1.
Industrial validation — Evaluated for compactness in video scene detection and for adaptation under distribution shift.

Paper Details

Length: 39 pages, 4 figures, 11 tables
Primary subject: Computer Vision and Pattern Recognition (cs.CV)
Secondary subject: Image and Video Processing (eess.IV)
Cite as: arXiv:2609.20869 [cs.CV]
Version: v1 (this version)

Takeaway

TAPe+ML v3 demonstrates that offloading modeling burden into a structured input representation can yield compact, multi-task vision systems with materially lower data, memory, and compute requirements—an increasingly important trade-off in 2026's deployment-focused landscape.

via ArXiv CV

TAPe+ML v3: A Compact Structured Representation for Multi-Task