SlotSAR: Unsupervised SAR Clutter Disentanglement

• The paper introduces SlotSAR, a novel unsupervised framework that disentangles target and background clutter in SAR imagery using dual-slot attention.
• It fuses high-level SARATR-X semantic features with low-level wavelet scattering tokens to balance spatial priors and structural detail for improved clutter suppression.
• Empirical results on the ATRNet-STAR benchmark show significant gains in ARI and mBO metrics, demonstrating robustness in complex and unseen operating conditions.

SlotSAR is an object-centric learning framework for synthetic aperture radar (SAR) imagery that is designed to disentangle target representations from background clutter without requiring pixel-level target masks during training. It reformulates feature-level clutter reduction as slot-based decomposition: instead of explicitly supervising a segmentation mask, it learns two latent slots, one intended to represent the target and the other the background clutter, and refines them iteratively so that downstream recognition can rely more on target-specific structure than on spurious background correlations. Its defining technical contribution is a multi-level slot attention mechanism that fuses high-level semantic features from a pretrained SAR foundation model with low-level scattering features from a wavelet scattering network, and its experimental evaluation is centered on the ATRNet-STAR benchmark under standard and extended operating conditions (Jang et al., 11 Sep 2025).

1. Problem formulation and conceptual position

SlotSAR is motivated by a specific failure mode of SAR automatic target recognition: SAR images contain not only targets of interest but also terrain reflections, vegetation, man-made structures, and speckle noise, and this clutter can exhibit intensity and pattern statistics that resemble the target. Standard deep models can therefore learn entangled features in which target and clutter are mixed, or they can exploit clutter-specific shortcuts that degrade under unseen backgrounds. Prior feature-level clutter reduction methods often depend on precise target masks, but such annotations are difficult to obtain reliably for SAR. SlotSAR addresses this by learning object-centric, slot-wise target/clutter decomposition from image-level training rather than mask supervision (Jang et al., 11 Sep 2025).

Conceptually, the method is a SAR-specific instance of object-centric learning. It does not attempt dense scene parsing, dense 3D reconstruction, or semantic part decomposition. Instead, it learns exactly two slots: a target slot and a background slot. This makes the model specialized to single-target decomposition rather than general multi-object scenes. In the paper’s framing, this constitutes the first unsupervised target/clutter disentanglement framework for SAR, with “unsupervised” referring specifically to the absence of mask annotations during training rather than to the absence of pretrained components or evaluation labels (Jang et al., 11 Sep 2025).

The central design claim is that target-background disentanglement in SAR requires both semantic context and scattering-sensitive structural detail. High-level semantic features alone can be too coarse and can entangle target and clutter when their statistics are similar, while low-level scattering features alone lack sufficient semantic context. SlotSAR therefore fuses the two feature types inside slot attention, rather than only at the encoder level, so that slot competition is guided simultaneously by coarse spatial priors and fine structural cues. This is the basis on which the method performs target-background separation without mask annotations (Jang et al., 11 Sep 2025).

2. Architectural composition and feature representations

Architecturally, SlotSAR comprises a high-level visual encoder, a low-level scattering encoder, a multi-level slot attention module, and a decoder. The input SAR image is denoted

$I \in \mathbb{R}^{H \times W}.$

The high-level branch uses a pretrained SARATR-X encoder $g_v$ to produce semantic tokens

$$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$

SARATR-X is described as a SAR foundation model pretrained with self-supervised masked patch reconstruction, and its features are said to capture high-level semantics while also being refined with multi-scale gradient features to suppress speckle and emphasize contours and edges. Within SlotSAR, these features provide coarse spatial and semantic guidance for object-centric decomposition (Jang et al., 11 Sep 2025).

The low-level branch is a wavelet scattering network $g_s$ that extracts scattering structure:

$$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$

The branch uses a learnable 2D Morlet wavelet parameterization,

$$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$

where $R_\theta$ is the rotation matrix, $D_\gamma$ is an anisotropic dilation matrix, $u' = u_1 \cos \theta + u_2 \sin \theta$, and $\beta$ enforces the zero-mean condition. The parameters $g_v$0 (scale), $g_v$1 (orientation), $g_v$2 (frequency), and $g_v$3 (aspect ratio) are learned during training. This branch is motivated by the observation that wavelet scattering is well suited to SAR because it can preserve structural detail, characterize local scattering behavior, and suppress speckle (Jang et al., 11 Sep 2025).

After pooling, resizing, and flattening into token form, the scattering tokens are processed by a lightweight channel-mixing encoder and self-attention:

$g_v$4

The resulting $g_v$5 is the final low-level structural representation. Before slot attention, both branches are normalized and projected:

$g_v$6

The architecture therefore preserves a strong asymmetry between branches: SARATR-X contributes coarse semantic localization, whereas the wavelet scattering path contributes fine structural and scattering-sensitive detail (Jang et al., 11 Sep 2025).

3. Multi-level slot attention and iterative slot refinement

The multi-level slot attention (MLSA) module is the core of SlotSAR. It uses exactly two slots,

$g_v$7

with $g_v$8 and $g_v$9 denoting target and background slots. Initial slots are sampled from a learned Gaussian,

$$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$0

where $$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$1 are learnable. Visual features are projected to keys and values,

$$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$2

and at iteration $$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$3 the query is

$$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$4

The first stage obtains coarse top-down spatial information from the semantic branch by forming slot-attention logits

$$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$5

and token-to-slot assignments

$$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$6

This attention map is shifted to produce a multiplicative spatial prior,

$$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$7

where $$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$8 is the mean attention value (Jang et al., 11 Sep 2025).

The second stage injects scattering structure. The slot-conditioned fusion tensor is

$$F_v = g_v(I), \qquad F_v \in \mathbb{R}^{N \times D_{\text{feat}}}.$$9

where $g_s$0 denotes an outer product between slot-specific spatial maps and scattering tokens. This fusion map is used to refine the visual value features through element-wise multiplication, $g_s$1, so that target-relevant value dimensions receive stronger responses. The paper then applies a second normalization over tokens,

$g_s$2

and updates slots by a GRU:

$g_s$3

A final MLP update gives

$g_s$4

The paper notes a likely shape typo in the manuscript for this final expression; conceptually, however, the operation is standard slot refinement in slot-embedding space (Jang et al., 11 Sep 2025).

This design makes the refinement process explicitly hierarchical. Semantic features first determine where each slot should attend, while scattering features modulate what structural patterns are emphasized inside those regions. The result is not generic slot attention transferred unchanged to SAR, but a SAR-specific top-down/bottom-up fusion mechanism. A common misconception is that SlotSAR discovers a full set of scene objects or semantic parts; it does not. Its object-centricity is restricted to a binary decomposition between target and clutter, with iterative refinement used primarily to sharpen target localization and suppress clutter leakage (Jang et al., 11 Sep 2025).

4. Optimization protocol and evaluation methodology

Training uses a lightweight MLP decoder $g_s$5 and a feature reconstruction objective. The decoder reconstructs visual encoder features from the refined target and background slots jointly, and the loss is an $g_s$6 reconstruction loss between the visual encoder features and the decoder output. The method follows the DINOSAUR style of training on pretrained features rather than direct pixel reconstruction. The paper does not define additional classification, segmentation, KL, entropy, or diversity regularization losses, and it does not present a classification head for the object-centric module itself. Binary segmentation masks are used only for evaluation (Jang et al., 11 Sep 2025).

The training procedure is described as two-stage. First, the model is pretrained on the training split of the EOC-scene dataset, which has relatively simple clutter, to stabilize object-centric representation learning. It is then fine-tuned on each scenario-specific dataset, and the “entire model” is fine-tuned. The paper states that training uses Adam with learning rate $g_s$7, linear warmup for 10k steps, exponential decay, gradient clipping at 1, batch size 64, and 200 epochs. The implementation uses an MLP decoder, slot dimension $g_s$8, $g_s$9 slot refinement iterations, scattering token dimension $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$0, $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$1 tokens, and semantic feature dimension $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$2. Experiments were run on a single RTX 4090 GPU with 24 GB memory (Jang et al., 11 Sep 2025).

Evaluation is centered on ATRNet-STAR, described as a large-scale SAR benchmark with 40 vehicle classes and 194,324 annotated images collected under complex operating conditions, including variations in band, polarization, scene, azimuth, depression, and target position. The paper evaluates under standard operating conditions and multiple extended operating conditions, including EOC-Azi., EOC-Band, EOC-Dep., EOC-Pol., and EOC-Scene. Although SlotSAR is trained without mask supervision, the dataset provides bounding boxes, and the authors build a tool using prior work to generate binary segmentation maps for quantitative evaluation (Jang et al., 11 Sep 2025).

The evaluation metrics are adjusted Rand index (ARI), mean best overlap (mBO), and mean intersection over union (mIoU). ARI measures clustering consistency between predicted slot masks and ground-truth segmentation. mBO measures the best overlap with the target region excluding background. mIoU measures average overlap over target and background. For quantitative comparison, decoder-generated alpha masks are matched to ground truth using Hungarian matching for mBO and mIoU. Baselines include Slot Attention, DINOSAUR, ISA, and SMSA, with DINOv2 and SARATR-X used as visual backbones when applicable and a shared decoder for fairness (Jang et al., 11 Sep 2025).

5. Empirical performance and ablation structure

Across 21 evaluation settings on ATRNet-STAR, SlotSAR achieves the best average ARI and mBO. The reported average results are ARI $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$3 and mBO $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$4, compared with the best baseline, SMSA with SARATR-X, at ARI $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$5 and mBO $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$6, corresponding to gains of $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$7 ARI and $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$8 mBO. The paper emphasizes especially difficult unseen-scene settings, where SlotSAR is reported as more robust to clutter conditions more complex than those seen in the pretraining stage (Jang et al., 11 Sep 2025).

Setting	SlotSAR	Comparison reported
Average over 21 ATRNet-STAR settings	ARI $$F_s = g_s(I), \qquad F_s \in \mathbb{R}^{C \times H' \times W'}.$$9, mBO $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$0	SMSA+SARATR-X: ARI $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$1, mBO $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$2
EOC-Scene overall	ARI $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$3, mBO $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$4	SMSA: ARI $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$5, mBO $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$6
Woodland	ARI $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$7, mBO $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$8	DINOSAUR+SARATRX: ARI $$\psi_{\sigma, \theta, \xi, \gamma}(u) = e^{\left(-\frac{ \| D_\gamma R_\theta(u) \|^2 }{2\sigma^2} \right)} \left( e^{(i \xi u') - \beta} \right),$$9, mBO $R_\theta$0
Factory	ARI $R_\theta$1, mBO $R_\theta$2	DINOSAUR+SARATRX: ARI $R_\theta$3, mBO $R_\theta$4

The ablation studies isolate the contributions of top-down spatial guidance, scattering features, and iterative slot refinement. On EOC-Scene, the paper reports: DINOSAUR baseline at ARI $R_\theta$5, mBO $R_\theta$6, mIoU $R_\theta$7; a VQ + $R_\theta$8 comparative baseline at ARI $R_\theta$9, mBO $D_\gamma$0, mIoU $D_\gamma$1; only spatial map $D_\gamma$2 at ARI $D_\gamma$3, mBO $D_\gamma$4, mIoU $D_\gamma$5; only scattering feature $D_\gamma$6 at ARI $D_\gamma$7, mBO $D_\gamma$8, mIoU $D_\gamma$9; and full SlotSAR with both $u' = u_1 \cos \theta + u_2 \sin \theta$0 and $u' = u_1 \cos \theta + u_2 \sin \theta$1 at ARI $u' = u_1 \cos \theta + u_2 \sin \theta$2, mBO $u' = u_1 \cos \theta + u_2 \sin \theta$3, mIoU $u' = u_1 \cos \theta + u_2 \sin \theta$4. These results support three claims made in the paper: vector quantization is harmful in this SAR setting because target and clutter are too similar; scattering features alone provide a stronger gain than spatial guidance alone; and combining both yields the best performance (Jang et al., 11 Sep 2025).

A second ablation studies the number of slot refinement iterations. On EOC-Scene, the baseline and SlotSAR results are reported as follows: at $u' = u_1 \cos \theta + u_2 \sin \theta$5, ARI $u' = u_1 \cos \theta + u_2 \sin \theta$6 vs $u' = u_1 \cos \theta + u_2 \sin \theta$7 and mBO $u' = u_1 \cos \theta + u_2 \sin \theta$8 vs $u' = u_1 \cos \theta + u_2 \sin \theta$9; at $\beta$0, ARI $\beta$1 vs $\beta$2 and mBO $\beta$3 vs $\beta$4; at $\beta$5, ARI $\beta$6 vs $\beta$7 and mBO $\beta$8 vs $\beta$9. The paper interprets this as progressive exploitation of the fusion map $g_v$00: at $g_v$01 the model already roughly localizes the target, and later iterations sharpen structure and suppress clutter more effectively (Jang et al., 11 Sep 2025).

Qualitative evidence is also part of the empirical argument. The visualizations are said to show that SlotSAR preserves finer target structure and is more resistant to clutter-induced entanglement under both natural clutter and artificial clutter. Under artificial clutter with stronger reflections than the target, baseline slot methods often attend to high-backscatter distractors, whereas SlotSAR more reliably isolates the target. The paper also includes t-SNE visualizations of slot embeddings, showing substantial overlap between target and background slot distributions for SMSA and much clearer separation for SlotSAR. This is presented as direct evidence that scattering-based structural guidance improves slot distinctiveness in SAR (Jang et al., 11 Sep 2025).

6. Scope, limitations, and relation to adjacent SAR representation learning

SlotSAR has a deliberately narrow formulation. It assumes two slots only, one target and one background, so it is specialized to single-target decomposition rather than general multi-object scenes. Several equations in the manuscript appear to have notation or shape typos, which means implementation requires consistent tensor-dimension resolution. The method also relies on a pretrained SARATR-X backbone and on a two-stage training curriculum, and the paper notes that backbone-update details during fine-tuning are not fully clarified. Finally, segmentation quality is evaluated indirectly using generated binary masks derived from dataset annotations rather than native segmentation labels (Jang et al., 11 Sep 2025).

These limitations clarify what SlotSAR should not be conflated with. It is not a part-based semantic model of the kind used in structural reconstruction, and it is not a generic scene decomposition model. A related but distinct direction is SAR2Struct, which defines SAR target structure recovery as inferring the components of a target and the structural relationships between its components, specifically symmetry and adjacency, from a single-view SAR image; its 3D semantic representation uses oriented bounding boxes and Symmetry Hierarchy, and therefore operates at the level of components and relations rather than target/background slots (Yue et al., 7 Jun 2025).

SlotSAR is also part of a broader trend toward scattering-aware SAR representation learning. The paper’s use of a wavelet scattering network is consistent with earlier work arguing that analytic scattering transforms provide noise-resilient and interpretable SAR features under speckle. For example, SSN proposes a Stockwell scattering network for SAR image change detection and explicitly frames scattering representations as noise-resilient, interpretable, and computationally efficient (Chen et al., 2023). SlotSAR differs in that scattering features are not the entire representation but one branch in a multi-level slot-attention system.

A plausible implication is that future slot-based SAR models could combine SlotSAR’s target/clutter disentanglement with richer structural priors. One direction would be to move from binary target/background decomposition toward component-level object-centricity; another would be to preserve SlotSAR’s mask-free training regime while introducing semantic relations analogous to symmetry, adjacency, or hierarchical part organization. The existing SlotSAR formulation does not attempt these extensions, but its emphasis on slot-wise structural distinctiveness suggests a natural bridge between clutter suppression and more semantic object-centric SAR representations.