TAPe+ML: A Compact Structured Representation for Multi-Task Computer Vision
Sergey Kurinov, Alexey Upatov
Comexp Research Lab
Published September 15, 2026 · on the feed September 22, 2026 · ▲ 2 · arXiv 2609.20869
The abstract
We present TAPe+ML v3, a compact computer vision system based on TAPe (Theory of Active Perception), a structured representation that encodes relations among perceptual elements before recognition. Instead of operating directly on pixel tensors, the system uses a shared TAPe representation and a modular recognition architecture for 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 obtains 84.7 mAP50 and 65.3 mAP50-95. On COCO instance segmentation, it obtains 80.7 mask mAP50 and 58.4 mask mAP50-95. In classification experiments, it reaches 92 percent validation accuracy on Imagenette under an identical-training comparison with a raw-pixel baseline, and 89.9 percent Top-1 accuracy on ImageNet-Real. We also evaluate compactness in video scene detection and adaptation under distribution shift in an industrial pilot. The 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.
The Brief
2:13A short podcast summary of the paper. Written and voiced by AI.
Transcript
TAPe+ML. A Compact Structured Representation for Multi-Task Computer Vision. By Sergey Kurinov and one other, from Comexp Research Lab. Published on the fifteenth of September. No code repository is linked to it.
Most vision systems start from pixel tensors and ask the network's parameters to work out everything else: what the parts of an image are, how they relate, and what they add up to. That puts the whole modeling burden on the network, and the network's size, data and compute grow to carry it.
TAPe+ML version three moves part of that burden into the input. It is built on TAPe, short for Theory of Active Perception, a structured representation that encodes relations among perceptual elements before recognition happens. The system does not operate on raw pixels. It uses one shared TAPe representation, with a modular recognition architecture on top, and serves three tasks from it: image classification, object detection, and instance segmentation.
Four pieces make up the architecture. There is background and contour processing. There is local object localization. Classification is prototype based. And a coordinator manages specialized submodels. Across all the reported experiments, the whole system uses fewer than one hundred thousand parameters.
On COCO object detection, it scores eighty-four point seven on mAP fifty, and sixty-five point three on mAP fifty to ninety-five. On COCO instance segmentation, the mask scores are eighty point seven on mAP fifty, and fifty-eight point four on mAP fifty to ninety-five.
For classification, it reaches ninety-two percent validation accuracy on Imagenette, in a comparison against a raw-pixel baseline trained identically. On ImageNet-Real, top-one accuracy is eighty-nine point nine percent. Beyond those, the authors evaluate compactness in video scene detection, and adaptation under distribution shift in an industrial pilot.
In the authors' words, the 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 requirements for data, memory and compute. The Imagenette comparison is the one set up to isolate that idea, since training is held identical and the input representation is what changes.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Most vision systems start from pixels and pay for it in size. This one asks why.
Your company runs one model for classification, another for detection, a third for segmentation. Each has its own data, formats and infrastructure, and the big backbones are pretrained on unrelated tasks just to work at all.
The road ahead
- 1 · The problem ~7 min Vision in industry runs as a zoo of separate pixel models, and each one must rebuild an image's structure from raw numbers before doing its real job.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Encode how image elements relate before recognition, so tiny modules sharing that one representation can handle several vision tasks under a coordinator.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min The image becomes a shared TAPe representation, then small modules separate background and contours, locate objects, and classify each by matching class prototypes.
- 6 · The experiment ~6 min You’ll guess: With a tiny parameter budget, what mAP50 does TAPe+ML v3 reach on COCO object detection?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.