Atomizer-IO: Beyond Pixels, Patches and Grids

2026-09-30

AI Summary

This study addresses the limitation of existing vision architectures that rely on regular grid assumptions, rendering them ineffective for unstructured sensor data. To overcome this, we propose a novel architecture based on atomic representations, introducing an "observation-first" paradigm. By leveraging measurement metadata and anchor-based local cross-attention mechanisms, our approach infers structural information directly from physical relationships, thereby eliminating grid constraints entirely. The proposed architecture seamlessly generalizes to diverse inputs, such as unordered 3D point clouds, without requiring task-specific redesigns. Furthermore, it achieves performance levels comparable to specialized models across multiple tasks while enabling post-training control over inference costs. These results comprehensively validate the versatility and flexibility of the proposed grid-free interface for processing heterogeneous sensory data.

Abstract

Most vision architectures assume that observations lie on a regular grid, an effective abstraction for natural images but a restrictive one for sensing data whose channels, temporal sampling, spatial resolution, and geometry can vary. Generic set-based architectures remove the grid, but also remove useful spatial inductive biases. We introduce Atomizer-IO, an architecture that places observations first and derives structure from their physical relationships. Building on top of an atomic representation of the data, each observation is described by its measurement and acquisition metadata, while local cross-attention maps observations to anchor points that can be arbitrarily placed. We evaluate this design by progressively relaxing the grid assumption, from varying input raster configurations and incomplete channel sets to flexible output density and, ultimately, inputs without a raster grid. Atomizer-IO is competitive with flexible EO-specific architectures on most tasks, while offering post-training control over inference cost and competitive compute--performance trade-offs. The same formulation extends without architectural redesign to unordered 3D point clouds, showing that the atomic interface generalizes beyond regular raster inputs. These results suggest that pixels, patches, and grids do not need to define the interface of a sensing architecture.

Problem

• vision architecture
• sensing data
• grid assumption
• spatial inductive biases
• atomic representation

Innovation

• Atomizer-IO
• atomic representation
• local cross-attention
• grid-free architecture
• 3D point clouds