Atomizer-IO: Beyond Pixels, Patches and Grids

Hugo Riffaud de Turckheim; Sylvain Lobry; Nicolas Houdré; Damien Robert; Roberto Interdonato; Diego Marcos

cs.CVSep 30, 2026

Atomizer-IO: Beyond Pixels, Patches and Grids

Authors: Hugo Riffaud de Turckheim, Sylvain Lobry, Nicolas Houdré, Damien Robert, Roberto Interdonato, Diego Marcos

Organizations: INRIA, Montpellier, France · LIPADE, Université Paris Cité, Paris, France · DM3L, University of Zurich, Zurich, Switzerland · CIRAD, Montpellier, France

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.

Figures & tables

Figure 1: Sensing configurations considered in this work. All are processed through the same atomic interface, without rasterization, resampling, or modality-specific branches.

Figure 2: The Atomizer-IO architecture. Atoms are locally aggregated into a compact spatial representation, processed globally in latent space, and decoded through spatial output queries. The same local context-to-query operation is used for encoding and decoding, with spatial relationships defined in physical coordinates rather than by a fixed input grid.

Figure 3: Token construction in Atomizer-IO. 2D imagery and LiDAR share the same atomic interface of measurements and acquisition metadata; spatial latents, global latents, and output queries use learned initializations.

Figure 4: Spatial sampler. From left to right: query and context sets; construction of a local receptive field from spatial proximity, illustrated by the encoder’s nearest-latent assignment and resulting Voronoi partition; query-centered relative coordinates; local cross-attention.

Figure 5: Example spatial latent partition on DALES (Section 4.4 ). Colors show the local receptive fields induced by assigning each input atom to its nearest spatial latent.

Dataset	Sensor	T	C	Geometry	GSD	Problem	Out. C
ForestNet ( Irvin et al., 2020 )	Landsat-8	1	6	$332^{2}$	15 m	cls	12
BurnScars ( Jakubik et al., 2023 )	HLS	1	6	$512^{2}$	30 m	seg	2
EuroSAT ( Helber et al., 2019 )	Sentinel-2	1	13	$64^{2}$	10 m	cls	10
Cashew ( Jin et al., 2021 )	Sentinel-2	1	13	$256^{2}$	10 m	seg	7
Sen1Floods11 ( Rambour et al., 2020 )	S1 + S2	1	15	$512^{2}$	10 m	seg	2
xView2 ( Gupta et al., 2019 )	RGB (VHR)	2	3	$512^{2}$	0.5 m	seg	5

Table 1: Datasets used for evaluation, spanning raster Earth observation and irregular LiDAR inputs. T : timesteps; C : input channels; Geometry : spatial structure of the input; GSD : ground sampling distance, undefined for point clouds; Out. C : output classes or regression targets.

Model	EuroSAT	ForestNet	BurnScars	Sen1Floods	PASTIS	xView2	Cashew	BioMassters $\downarrow$
ResNet50	90.9	43.8	82.8	87.8	30.0	53.8	69.9	46.3
ViT	90.7	39.1	87.5	85.0	35.1	52.4	47.5	52.7
RAMEN	92.1	40.4	88.2	89.1	45.5	53.3	64.2	44.7
UniverSat	90.8	44.3	88.1	86.8	42.9	56.3	62.4	46.0
Perceiver-IO	90.9	35.4	83.4	92.2	17.9	43.8	26.2	52.2
Atomizer-IO	90.2	35.8	88.8	93.2	44.1	54.6	72.0	42.4

Table 2: Single-task performance. Best in bold , second best underlined . Classification uses macro F1, segmentation mIoU, and BioMassters RMSE (Mg/ha,

\downarrow

	Sen1Floods11 $\uparrow$						BioMassters $\downarrow$
Test Configuration	Atom.	Res.	ViT	Perc.	RAM.	Uni.	Atom.	Res.	ViT	Perc.	RAM.	Uni.
All bands	92.9	88.1	84.5	92.1	87.0	88.7	41.5	47.6	52.9	52.0	43.4	45.7
S2 only	92.8	86.4	84.4	92.2	87.8	88.6	45.8	55.9	55.5	57.7	48.4	49.1
S1 only	80.0	66.4	75.3	77.2	75.1	79.4	48.1	52.1	55.5	58.6	49.2	49.1
RGB only	44.8	43.7	43.8	43.8	44.5	43.8	71.0	73.8	71.4	70.4	71.9	70.8
No SWIR	87.8	49.0	75.1	81.8	63.7	85.2	43.0	49.1	54.0	53.4	46.0	46.7

Table 3: Performance under flexible input composition on Sen1Floods11 (mIoU, %) and BioMassters (RMSE, Mg/ha). All models are trained under the same band-dropout protocol.

Figure 6: Compute cost (GFLOPs) versus segmentation performance (mIoU). Dense, Zone-probe, and Quadtree decoding vary inference cost through different output query strategies, using the same Atomizer-IO checkpoint without retraining.

DALES
Method	mIoU
KPConv ( Thomas et al., 2019 )	81.1
SPT ( Robert et al., 2023 )	79.6
PointNet++ ( Qi et al., 2017b )	68.3
ConvPoint ( Boulch, 2020 )	67.4
Atomizer-IO	66.7

Table 4: DALES and FRACTAL segmentation results (mIoU, %).

	MNIST	Sen1Fl11
No RoPE	29.48	92.42
RoPE	99.37	93.17

Table 5: RoPE ablation. Accuracy on MNIST and mIoU on Sen1Floods11.

Background kept (%)	Atomizer-IO	ViT
$100$	99.26	98.11
$75$	98.72	97.65
$50$	98.37	94.85
$25$	94.93	72.17
$10$	74.36	37.24

Table 6: MNIST accuracy (%) under unseen input geometries. Digit pixels are always retained; both models use the same checkpoint trained on complete inputs.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Detailed Atomizer-IO architecture. The encoder aggregates local neighborhoods of input atoms into spatial latents through the spatial sampler; the processor applies latent self-attention with optional temporal aggregation; and the decoder maps local latent neighborhoods to spatial output queries, optionally conditioning first on raw observations at the target location. Relative positional information is introduced locally, and MLP and attention weights are shared across spatial neighborhoods.

Figure 8: Temporal aggregation ablation on PASTIS as the number of input timesteps increases.

	MNIST	Cashew	PASTIS	BurnScars	Sen1Floods11
No RoPE	29.48	62.65	40.65	86.16	92.42
RoPE	99.37	72.01	44.10	88.81	93.17

Appendix

Table 7: Effect of spatial RoPE across tasks. MNIST reports accuracy (%), while the EO datasets report mIoU (%).

Figure 9: Relative degradation when spatial RoPE is removed, computed as

100(\mathrm{RoPE}-\mathrm{No~RoPE})/\mathrm{RoPE}

	Sen1Floods11	Cashew	PASTIS	BurnScars
Without local observations	89.26	70.58	43.30	88.20
Full model	93.17	72.01	44.10	88.81

Appendix

Table 8: Effect of conditioning output queries on the raw observations at their target location. All values are mIoU (%).

Figure 10: Relative degradation when local observation conditioning is removed, computed as

100(\mathrm{Full}-\mathrm{No~Local})/\mathrm{Full}

	Interaction bound	Equiv. global drop	Context/query
Global	$LN\approx 7.86\times 10^{9}$	0%	$N$
Local ( $m{=}500$ )	$\leq Lm=1.0\times 10^{6}$	99.987%	$\leq 500$
Local ( $m{=}1000$ )	$\leq Lm=2.0\times 10^{6}$	99.975%	$\leq 1000$
Local ( $m{=}2000$ )	$\leq\min(N,Lm)\approx 3.93\times 10^{6}$	99.949%	$\leq 2000$

Appendix

Table 9: Cross-attention complexity for global and local spatial sampling on Sen1Floods11 (

N\approx 3.9\times 10^{6}

atoms,

L=2000

spatial latents). The equivalent global drop rate is

1-m/N

, the fraction of atoms that global cross-attention would need to discard to match a budget of

m

context elements per latent.

	Full decode calls	Best case	Worst case
Dense	$M$	$M$	$M$
Zone-probe	$k_{p}L+\sum_{\ell\in\mathrm{hard}}\|W_{\ell}\|$	$\mathcal{O}(k_{p}L)$	$M$
Quadtree	$\lesssim k_{p}\times(\text{levels}\times\text{active cells})$	$\mathcal{O}\!\left(\dfrac{k_{p}HW}{s_{\text{start}}^{2}}\right)$	$M$

Appendix

Table 10: Decode cost of the three inference strategies.

M

is the number of output queries,

L

the number of spatial latents,

k_{p}

the number of probes per region, and

s_{\text{start}}

the initial quadtree cell size. Both adaptive strategies reduce to dense decoding in the worst case when no region can be treated as homogeneous.

Figure 11: Sensitivity of segmentation performance and inference cost to three architectural parameters: atoms per latent, encoder cross-attention budget, and decoder neighborhood size. Results are shown on Sen1Floods11 and BurnScars for Dense, Zone-probe, and Quadtree decoding.

Figure 12: LiDAR return encoding. The pair

(a,b)

describes the relative position of each observation within its pulse’s return sequence. Normalization by the per-pulse return count

R

makes the representation independent of the sensor-specific maximum number of returns while preserving the original

(r,R)

attributes.

Architecture	Classification	Segmentation	Multitemporal seg.
ResNet50	23.6	37.3	37.3
ViT-Small	21.7	30.3	31.5
Perceiver IO	19.7	34.5	34.5
RAMEN	23.1	33.7	33.7
UniverSat	36.1	36.1	36.1
Atomizer-IO	19.2	33.5	34.8

Appendix

Table 11: Number of trainable parameters for the architectures used in the 2D experiments. Values are given in millions of parameters. ViT refers to ViT-Small.

	RAMEN	UniverSat
Dataset	Enc. res. (m)	Enc. res. (m)	Subpatch (px)	Dec. stride (px)
ForestNet	40	40	1	–
BurnScars	120	120	1	4
EuroSAT	40	40	1	–
Cashew	40	40	1	1
Sen1Floods11	40	40	1	4

Appendix

Table 12: Architecture-specific spatial hyperparameters used for RAMEN and UniverSat. Encoding resolutions are expressed in physical units.

Dataset	Batch size
ForestNet	8
BurnScars	8
EuroSAT	8
Cashew	8
Sen1Floods11	4
xView2	8

Appendix

Table 13: Batch size used for each dataset. The same batch size is used across architectures within a dataset.

Hyperparameter	2D	3D
Latent dimension	512	768
Global latents	128	128
Attention heads	8	8
Decoder neighbors	9	9
Latent layout	Hexagonal	Hexagonal
$(k,m)$	Other 2D: $\{(1000,1000),(1500,1500),(2000,2000)\}$ PASTIS: $\{(100,100),(250,250),(350,350)\}$	$(1000,1000)$

Appendix

Table 14: Main Atomizer-IO architectural hyperparameters used for the 2D and 3D experiments. Here,

k

denotes the target number of atoms per spatial latent and

m

the maximum number of atoms sampled per latent for local cross-attention.

Feature	Frequency bands	Maximum frequency / period
Position	16	16
Resolution	2	2
Acquisition time	12	365 days
Measurement value	8	8

Appendix

Table 15: Fourier feature parameters used by Atomizer-IO and Perceiver IO.

Figure 13: MNIST sparsification protocol. Models are trained on complete images. At evaluation, pixels with value $v\leq 0.5$ are progressively removed, while pixels with value $v>0.5$ are always retained.

Parameter	Atomizer-IO	ViT
Parameters (total)	7.40M	7.44M
Hidden / latent dimension	256	224
Encoder / processor depth	1 cross-attn + 4 self-attn	12 transformer blocks
Attention heads	4 (cross) / 8 (self)	4
MLP dimension	768	896 ( $4\times$ )
Positional encoding	Relative (RoPE, self-attn)	Absolute (learned)

Appendix

Table 16: Architecture and training hyperparameters for the capacity-matched Atomizer-IO and ViT models used in the MNIST sparsification experiment.

Method	Mean	Ground	Buildings	Cars	Trucks	Poles	Power lines	Fences	Vegetation
KPConv	0.811	0.971	0.966	0.853	0.419	0.750	0.955	0.635	0.941
PointNet++	0.683	0.941	0.891	0.754	0.303	0.400	0.799	0.462	0.912
ConvPoint	0.674	0.969	0.963	0.755	0.217	0.403	0.867	0.296	0.919
Atomizer-IO	0.667	0.962	0.941	0.670	0.095	0.549	0.829	0.395	0.895
SuperPoint	0.606	0.947	0.934	0.629	0.187	0.285	0.652	0.336	0.879
PointCNN	0.584	0.975	0.957	0.406	0.048	0.576	0.267	0.526	0.917

Appendix

Table 17: Per-class IoU on DALES. Values are reported as fractions. Baseline values are those reported for the corresponding methods on DALES.

Stage	Symbol	Meaning
Input	$N$	Number of input atoms
	$v_{i}$	Measured value of atom $i$
	$J$	Number of metadata fields of an atom
	$u_{i1},\dots,u_{iJ}$	Metadata fields of atom $i$
	$\phi_{0}(\cdot)$	Encoder of the measured value
	$\phi_{1}(\cdot),\dots,\phi_{J}(\cdot)$	Encoders of the metadata fields

Appendix

Table 18: Notation used throughout the methodology.

Explore similar work

Jun 22, 2026cs.CV

UniverSat: Resolution- and Modality-Agnostic Transformers for Earth Observation

Vision Transformers (ViT) dominate computer vision. However, their reliance on rigid patch projectors hinders transfer to Earth Observation (EO), where input modalities, scales, and resolutions vary widely. We introduce UniverSat, a ViT-style backbone built around a Universal Patch Encoder that maps patches from arbitrary spatial, spectral, and temporal resolutions, and from both optical and non-optical sensors, into a shared embedding space with a shared set of weights. This enables training a single model on heterogeneous multimodal corpora via self-supervision, yielding robust, sensor-agnostic spatial features. We validate this approach with strong results across classification and segmentation on standard EO benchmarks from GeoBench, PANGEABench, and SpectralEarth. Our code and models are available at https://github.com/gastruc/UniverSat.

Yohann Perron, Guillaume Astruc, Nicolas Gonthier +2

LIGM, Ecole Nationale des Ponts et Chaussées, IP Paris, Univ Gustave Eiffel, CNRS · EFEO · LASTIG, Univ Gustave Eiffel, IGN, ENSG +2

Trustworthy Earth Observation Decisions Earth Observation Self-Supervised Vision Transformers Detection Transformer Vision Transformer Transformer Architectures+1

Ask PDF

Jun 12, 2026cs.CV

RATS! Patches Talk Through Registers: Emergent Parts in Register Attention Transformers

When humans see a bird, they recognize far more than just "bird" -- they see a head, wings, and talons, a structured assembly of reusable parts that can be identified across every bird they have ever seen. We ask whether a self-supervised visual model can discover the same compositional structure on its own. To this end, we propose RATS (Register Attention Transformers), which decomposes the classification token into N learnable register tokens that route patch information through an L->N->N->L bottleneck via a three-step compress-communicate-broadcast attention. The N registers are partitioned across the H attention heads, so that registers assigned to different heads do not interact with each other. Without auxiliary losses or part annotations, each register spontaneously specializes into a proto-semantic region whose emerging structure resembles object parts. RATS surpasses all baselines by +12 mIoU on average across five segmentation benchmarks, with consistent gains on ADE20K (+1.11 mIoU) and COCO (+0.2 AP^m). Its register dictionary further exhibits part-level consistency and semantic proximity across related categories. Our results suggest that RATS may provide a useful architectural prior for structured and interpretable visual representation learning.

Timing Yang, Predrag Neskovic, Jansen Seheult +4

1Johns Hopkins University · 2Office of Naval Research, Arlington, VA · Department of Laboratory Medicine and Pathology, Mayo Clinic,FLOPs (G)MN, USA

Self-Supervised Vision Transformers Transformer Attention Bird Emergence

Ask PDF

Mar 3, 2026cs.CV

Utonia: Toward One Encoder for All Point Clouds

We dream of a future where point clouds from all domains can come together to shape a single model that benefits them all. Toward this goal, we present Utonia, a first step toward training a single self-supervised point transformer encoder across diverse domains, spanning remote sensing, outdoor LiDAR, indoor RGB-D sequences, object-centric CAD models, and point clouds lifted from RGB-only videos. Despite their distinct sensing geometries, densities, and priors, Utonia learns a consistent representation space that transfers across domains. This unification improves perception capability while revealing intriguing emergent behaviors that arise only when domains are trained jointly. Beyond perception, we observe that Utonia representations can also benefit embodied and multimodal reasoning: conditioning vision-language-action policies on Utonia features improves robotic manipulation, and integrating them into vision-language models yields gains on spatial reasoning. We hope Utonia can serve as a step toward foundation models for sparse 3D data, and support downstream applications in AR/VR, robotics, and autonomous driving.

Yujia Zhang, Xiaoyang Wu, Yunhan Yang +6

The University of Hong Kong · Xiaomi · The Chinese University of Hong Kong

Point Clouds Spatial Reasoning Encoders Autonomous Driving

Ask PDF