acceptodds
Under review as a conference paper at ICLR 2027

SlotODE: Slot Trajectory Separation for Collision Reduction

Abstract

Object-Centric Learning (OCL) aims to decompose a scene into slot representations corresponding to individual objects, and Slot Attention is a representative framework that iteratively updates these slots. However, how slot representations evolve across updates and whether they maintain separation remain unclear. For clarity, we define slot-state collision as the case in which initially different slot representations become identical during iterative updates, so that the collided slots can no longer maintain different representations for different objects. To address this problem, we propose SlotODE, which formulates iterative slot updates as an ordinary differential equation (ODE). Instead of directly predicting the next slot representations, SlotODE learns a vector field that defines the time derivatives of the slot representations. Our theoretical analysis shows that SlotODE prevents slot-state collision over any finite update interval, and further derives an explicit lower bound on pairwise slot separation. In experiments, SlotODE significantly improves object discovery performance across OCL benchmarks. In-depth analyses demonstrate the impact of the ODE formulation and its components, and the robustness of the learned vector field across different ODE solvers and numbers of update steps.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Object-Centric Learning, Representation Learning, Ordinary Differential Equation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_S1d3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SlotODE, an ODE-based iterative slot update mechanism for unsupervised object-centric learning. The authors introduce the concept of slot-state collision, where initially distinct slot representations become identical during iterative updates. To address this issue, SlotODE models slot updates as continuous-time dynamics and theoretically guarantees the exclusion of finite-time slot-state collisions under certain assumptions. Experimental results demonstrate improved object discovery performance across multiple benchmarks, including MOVi-C/E, COCO, and CLEVR10.

Critical strengths
  1. The proposed method consistently outperforms conventional Slot Attention on MOVi-C, MOVi-E, COCO, and CLEVR10. These improvements across different datasets and training settings provide meaningful evidence of the practical effectiveness of the proposed update mechanism.

  2. The paper provides a clear mathematical formulation of slot-state collision and establishes theoretical guarantees on finite-time collision exclusion and pairwise slot separation. This provides an interesting perspective on understanding iterative slot updates in object-centric learning.

Critical weaknesses

1. Empirical Validation of Slot-State Collision

The paper defines slot-state collision as the case where two initially distinct slot representations become exactly identical during iterative updates. However, it remains unclear how frequently such exact collisions occur in conventional Slot Attention (or even this really occurs) and whether they are a significant cause of object-level redundancy or fragmentation.

While the paper provides qualitative visualizations of slot trajectories, additional quantitative evidence would strengthen the motivation.

Here are my suggestions (just suggestions, not requirements, and alternative approaches are welcome if deemed more appropriate) :

  • Measuring pairwise slot distances throughout iterative updates and analyzing the frequency of exact or near-collisions could help establish whether this phenomenon occurs in practice.
  • Examining the relationship between slot-state collisions and object-level redundancy or fragmentation would help clarify whether preventing such collisions directly contributes to the observed performance improvements.

2. Hyperparameter Sensitivity & Heuristic Choices.

The method involves several hyperparameters that vary across datasets.The paper would benefit from further clarification of how these dataset-specific values were determined and whether the results are sensitive to these choices.

For example, why is the target retention ratio fixed at 0.5, and how sensitive is the model to this choice? A sensitivity analysis over different retention ratios would help establish the robustness of this heuristic. Furthermore, since the method is unsupervised, clarifying how these hyperparameters are selected without relying on ground-truth object annotations would strengthen the experimental evaluation.

Questions for authors

In Table 2, SlotODE achieves strong FG-ARI results, but SPOT with an autoregressive Transformer decoder achieves substantially higher mBO scores. Since SlotODE mainly modifies the slot update mechanism, have the authors considered using a Transformer or autoregressive decoder instead of the MLP decoder? Could this improve mBO and narrow the performance gap?

Additionally, as discussed in Weakness 1, it would be interesting to investigate whether slot-state collisions actually occur under different decoder architectures and whether the collision problem persists regardless of the decoder choice.

6 comments

We appreciate Reviewer S1d3 for the insightful comments and constructive feedback.

W1 Quantitative evidence of slot separation and mask overlap during Euler updates

To address both concerns, we compare Slot Attention and SlotODE on MOVi-C under the same adaptive slot-selection mechanism. Our two numerical analyses show that SlotODE better preserves the relative separation of the final closest slot pair and exhibits less mask overlap than Slot Attention, consistent with its better segmentation performance.

(1) Quantitative analysis of slot separation

We quantify slot proximity at the final update using the ratio of the minimum pairwise distance to the mean pairwise distance, and assess whether the closest pair’s predicted segmentation masks overlap.

For each image XnX_n and update step tt, we compute the Euclidean distance between every pair of selected slots:

dn,ij(t)=∥sn,i(t)−sn,j(t)∥2,(i,j)∈Cn,d_{n,ij}^{(t)}=\left\|s_{n,i}^{(t)} - s_{n,j}^{(t)}\right\|_2,\qquad (i,j)\in\mathcal{C}_n,

where Cn\mathcal{C}_n contains all unordered pairs of selected slots in image nn. We then calculate the mean pairwise distance and the minimum pairwise distance for each image:

dˉn(t)=1∣Cn∣∑(i,j)∈Cndn,ij(t),dn,min⁡(t)=min⁡(i,j)∈Cndn,ij(t).\bar{d}_{n}^{(t)} = \frac{1}{\vert \mathcal{C}_n \vert} \sum _{(i,j) \in \mathcal{C}_n} d _{n, ij}^{(t)} , \quad d _{n,\min} ^{(t)} = \min _{(i,j) \in \mathcal{C}_n} d _{n,ij}^{(t)}.

We average the mean and minimum pairwise distances across images as Dpair(T),Dmin(T)D_{\mathrm{pair}}^{(T)} , D_{\mathrm{min}}^{(T)}. At the final update, we report their ratio, Dmin(T)/Dpair(T)D_{\mathrm{min}}^{(T)}/D_{\mathrm{pair}}^{(T)}, to measure the closest-pair distance relative to the overall pairwise distance.

At the final update TT, we identify the pair attaining the minimum distance in each image and compute the IoU between their predicted segmentation masks. This measures whether the closest latent pair also has overlapping masks.

The results show the following:

  1. Slot Attention has a smaller minimum-to-mean distance ratio than SlotODE (0.19 vs. 0.57), indicating that Slot Attention’s closest pair is more compressed, whereas SlotODE’s is more separated, relative to each method’s mean pairwise distance. At the final update, Slot Attention has $D_{\mathrm{pair}}^{(T)}=$5.68 and $D_{\mathrm{min}}^{(T)}=$1.05, whereas SlotODE has $D_{\mathrm{pair}}^{(T)}=$7.68 and $D_{\mathrm{min}}^{(T)}=$4.35.
  2. The closest pair under Slot Attention also has greater overlap between its predicted masks than the closest pair under SlotODE (0.47 vs. 0.10). Together, these results indicate that SlotODE’s closest pair remains more distinct both in latent space and in its predicted masks.
Dmin(T)/Dpair(T)D_{\mathrm{min}}^{(T)} / D_{\mathrm{pair}}^{(T)} (↑\uparrow)IoU of closest pair (↓\downarrow)
Slot Attention0.190.47$_{\pm .02}$
SlotODE0.570.10$_{\pm .01}$

We have newly added Appendix B.5 to the revised manuscript. It summarizes the analysis above and reports Dpair(t)D_{\mathrm{pair}}^{(t)} and Dmin(t)D_{\mathrm{min}}^{(t)} , and their ratio at initialization and every update step.

(2) Slot separation preservation during updates

We also present an additional analysis that measures how the distance between the final closest pair changes from initialization to the final output. At each step, we measure its distance relative to the mean pairwise distance. The following results show stronger preservation of relative slot separation under SlotODE than under Slot Attention.

To measure how the final closest pair’s relative separation changes from initialization to the final update, we compute an image-wise relative separation change ratio RnR_n. For each image XnX_n, we identify the pair with the minimum distance at the final update and track this pair across all update steps. We normalize its distance at each step by the mean pairwise distance among the selected slots in that image:

(in∗,jn∗)=arg⁡min⁡(i,j)∈Cndn,ij(T),rn(t)=dn,in∗jn∗(t)dˉn(t),Rn=rn(T)rn(0).(i_n^\ast, j_n^\ast) =\arg \min_{(i,j)\in\mathcal{C}_n} d _{n,ij}^{(T)} , \qquad r_n^{(t)} = \frac{d _{n,i_n^\ast j_n^\ast}^{(t)}}{\bar d_n^{(t)}} , \qquad R_n=\frac{r_n^{(T)}}{r_n^{(0)}}.

Rn<1R_n<1 indicates that the pair’s distance has decreased relative to the other selected pairs, while Rn>1R_n>1 indicates that its relative distance has increased. We report the mean RnR_n, the fraction of images for which RnR_n falls below each threshold, and the mean IoU of the final masks for those images.

The results show the following:

  1. The mean RnR_n is below 1.0 under Slot Attention (0.65), indicating relative contraction of the final closest pair on average. Under SlotODE, it is above 1.0 (1.82), indicating relative expansion on average.
  2. While 5.41% of Slot Attention images have Rn<0.1R_n<0.1, no SlotODE images fall below this threshold. These results indicate that the final closest pair retains stronger relative separation under SlotODE.
  3. Among images with Rn<0.7R_n<0.7, the final closest pair’s mean mask IoU is lower under SlotODE than under Slot Attention (0.23 vs. 0.32). Slot Attention images with the more severe Rn<0.1R_n<0.1 have a mean IoU of 0.86, whereas SlotODE’s rare Rn<0.7R_n<0.7 cases have a mean IoU of 0.23. Stronger preservation of relative separation under SlotODE is accompanied by lower mask overlap.
MethodMean Rn(↑)R_n (\uparrow)
Slot Attention0.65
SlotODE1.82
MethodRn<0.1R_n<0.1 (%)(↓)(\downarrow)IoU (Rn<0.1)(↓)(R_n<0.1)(\downarrow)Rn<0.7R_n<0.7 (%)(↓)(\downarrow)IoU (Rn<0.7)(↓)(R_n<0.7) (\downarrow)
Slot Attention5.41%0.86$_{\pm .10}$56.99%0.32$_{\pm .16}$
SlotODE0%--0.58%0.23$_{\pm .15}$

We have newly added Appendix B.6 to the revised manuscript. It reports results at each update step and visualizes the image-wise distribution of RnR_n. We would appreciate the reviewer’s consideration of these additional results.

W2 Hyperparameter sensitivity

We agree that showing how sensitive the results are to these choices would help readers assess the method.

We use r0=0.5r_0=0.5 across all datasets as a simple, shared target that encourages the model to retain approximately half of the candidate slots in expectation. This value was not chosen from ground-truth object annotations or a dataset-specific estimate of object counts.

We agree, however, that the most suitable retention ratio may differ across datasets depending on scene complexity and object composition. To examine sensitivity to this choice, we varied r0r_0 on MOVi-C while keeping the other SlotODE settings fixed. We used two runs for each additional setting; the r0=0.5r_0=0.5 row reports the original result.

r0r_0FG-ARI ↑\uparrowmBO$_i$ ↑\uparrow
0.340.44$_{\pm 8.5}$15.21$_{\pm 5.8}$
0.457.48$_{\pm 3.3}$25.69$_{\pm 0.7}$
0.579.15$_{\pm 0.2}$38.62$_{\pm 0.1}$
0.678.63$_{\pm 0.2}$38.30$_{\pm 0.1}$
0.775.78$_{\pm 0.3}$36.87$_{\pm 0.2}$

The results are similar at r0=0.5r_0=0.5 and 0.60.6, but performance declines at 0.3 and 0.4. The comparable result at r0=0.6r_0=0.6 suggests that 0.5 is not a narrowly tuned choice on MOVi-C, while the decline at lower ratios shows that the choice of target still matters.

The improvement from SlotODE is also observed without the adaptive slot-selection loss, as shown in Appendix B.1 (Table 8). The autoregressive Transformer decoding results presented below likewise show an FG-ARI improvement without this loss. Thus, the observed benefit of the SlotODE update does not rely solely on the choice of r0r_0 and adaptive slot selection. Determining the most suitable number of slots for different scenes remains an important question in object-centric learning, and we have added this sensitivity analysis to Appendix C.3 of the revised manuscript, with the new text highlighted in magenta.

Q1 SlotODE with autoregressive decoder

During the rebuttal period, we conducted additional experiments with an autoregressive Transformer decoder (Auto.) on MOVi-C and COCO. We used a DINOSAUR setup with a pretrained ViT encoder and compared SlotODE with conventional Slot Attention under the same decoder. Each new configuration was run twice.

AdaSlot's adaptive slot-selection mechanism suppresses unselected slots through the alpha masks produced by a mixture decoder. Its direct application to segmentation masks derived from autoregressive decoder attention has not been established. We therefore conducted this comparison without adaptive slot selection for either method.

  • With the autoregressive decoder held fixed, SlotODE achieves higher FG-ARI but lower mBO$_i$ than Slot Attention on both MOVi-C and COCO.
  • Compared with SlotODE with MLP decoder, SlotODE with the autoregressive Transformer decoder achieves higher mBO$_i$ but lower FG-ARI on both datasets. Conventional Slot Attention also shows the same direction of change relative to its MLP-decoder results in Appendix B.1 (Table 8).
MethodDecoderAdaptive slot selectionDatasetFG-ARImBO$_i$
Slot Attention (new)Auto.XMOVi-C54.76$_{\pm 2.0}$45.93$_{\pm 1.0}$
SlotODE (new)Auto.XMOVi-C62.75$_{\pm 0.5}$42.89$_{\pm 0.2}$
SlotODE (reported)MLPXMOVi-C67.27$_{\pm 0.3}$36.51$_{\pm 0.4}$
Slot Attention (new)Auto.XCOCO31.99$_{\pm 0.6}$31.95$_{\pm 0.6}$
SlotODE (new)Auto.XCOCO33.62$_{\pm 0.1}$31.79$_{\pm 0.4}$
SlotODE (reported)MLPXCOCO40.81$_{\pm 0.1}$28.55$_{\pm 0.3}$

The results show the following:

These findings suggest that decoder choice can reduce the mBO$_i$ gap raised by the reviewer while retaining SlotODE’s slot update mechanism.

Q2 Collision in different decoders

We agree that decoder architecture can affect the slot states learned through reconstruction and may change how closely they approach one another during refinement. However, replacing the decoder does not guarantee that initially distinct slots remain distinct under conventional Slot Attention’s discrete update. SlotODE focuses on preserving slot separation through its update mechanism.

We additionally analyze slot separation with the autoregressive Transformer decoder (Auto.) on MOVi-C. Following W1, we compute Dmin(t)/Dpair(t)D_{\mathrm{min}}^{(t)}/D_{\mathrm{pair}}^{(t)} at each update step and report its final value together with the mask IoU of the closest pair. With the same autoregressive decoder, SlotODE has a higher final minimum-to-mean distance ratio (0.40 vs. 0.23) and lower closest-pair mask IoU (0.15 vs. 0.47) than Slot Attention.

Dmin(T)/Dpair(T)D_{\mathrm{min}}^{(T)} / D_{\mathrm{pair}}^{(T)} (↑\uparrow)IoU of closest pair (↓\downarrow)
Slot Attention0.230.47$_{\pm .05}$
SlotODE0.400.15$_{\pm .03}$

We examine whether the benefit of changing the slot update persists across decoder architectures.

  • Autoregressive Transformer (new, Q1): With the decoder held fixed, SlotODE achieves higher FG-ARI than Slot Attention on MOVi-C and COCO without adaptive slot selection.
  • CNN decoder: On CLEVR10, SlotODE also improves FG-ARI (96.64 vs. 94.72) and mBO$_i$ (22.45 vs. 21.04) over conventional Slot Attention (Table 1).

Together with the MLP-decoder analysis in W1, these measurements show that conventional Slot Attention produces close slot pairs with overlapping masks under both MLP and autoregressive Transformer decoders. SlotODE shows greater relative separation and lower mask overlap in both settings. The CLEVR10 result further shows that its performance benefit extends to a CNN decoder.

Reviewer_fBcZOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SlotODE, which replaces the recurrent updates in Slot Attention with a learned, time-conditioned ordinary differential equation over the joint slot state. Using permutation equivariance and uniqueness of ODE solutions, the authors prove that initially distinct slots cannot become identical within a finite time interval, and derive an exponential lower bound on their pairwise distances. The implementation combines numerical integration with spectral normalization, constrained Dynamic Tanh activations, input-dependent initialization, and adaptive slot selection. Experiments on MOVi-C, MOVi-E, COCO, and CLEVR10 report improvements over conventional Slot Attention.

Critical strengths

S1. The analysis of slot-state separation correctly treats the slots as a coupled system and uses a slot-swapped trajectory to establish collision exclusion. This provides a precise statement beyond the usual qualitative discussion of slot redundancy.

S2. Consistent reported improvements across several evaluation settings. These results provide meaningful evidence that the proposed update architecture is useful beyond a single benchmark or one particular slot-selection configuration.

Critical weaknesses

W1. The central separation property is not adequately connected to the practical failure mode or the implemented updates. The theory excludes exact equality of slot states, whereas redundant object assignments can occur between distinct states. Appendix A.6 explicitly acknowledges that state separation does not guarantee distinct object assignments and that the continuous-time guarantees do not automatically extend to numerical discretization. Furthermore, the main implementation uses three Euler steps, which do not automatically inherit the continuous-time guarantee. Nevertheless, the solver experiments measure segmentation performance rather than preservation of slot separation. Consequently, the evidence supports the effectiveness of the architecture more strongly than the paper’s proposed collision-based explanation.

W2. Experimental attribution and reporting require clarification before the comparisons can be considered conclusive. Several components change alongside the update formulation. The direct next-state ablation also removes the residual connection and time conditioning, limiting its ability to isolate the continuous-time formulation. Reported results also differ between tables: the SPOT variant with self-training and an MLP decoder on COCO has FG-ARI (41.95±0.8) in Table 2 but (40.35±2.73) in Table 22. Together, these issues limit the attribution of the gains to the proposed update formulation and reduce confidence in the consistency of the baseline comparisons.

Questions for authors

Q1. What evidence establishes that the three-step Euler implementation preserves meaningful slot separation? Either a condition verified for the implemented model or direct numerical measurements would help bridge the continuous-time analysis and the reported experiments.

Q2. Please reconcile the SPOT_MLP results in Tables 2 and 22, the MOVi-C baseline results in Tables 1 and 9–10, and the regularization descriptions in C.3, D.1, D.2, and Table 16. Do these refer to different experimental settings or reporting errors, and do the corrections change the principal comparisons?

Q3. Do you have measurements of pairwise slot distances in the original representation space, before and throughout updates, for both methods? In particular, does reduced near-collision frequency correspond to reduced redundant object assignments or fragmentation after accounting for representation scale?

7 comments

We thank Reviewer fBcZ for the constructive feedback and insightful suggestions to improve our work.

W1-1 & Q3 Slot separation and object assignments under numerical updates

To examine the practical behavior directly, we measure slot distances in the original representation space throughout the Euler updates on MOVi-C. We compare Slot Attention and SlotODE under the same adaptive slot-selection mechanism, and relate the closest slot pair to overlap between its predicted segmentation masks. Our two analyses show that SlotODE better preserves the relative separation of the final closest slot pair and exhibits less mask overlap than Slot Attention, consistent with its better segmentation performance.

(1) Quantitative analysis of slot separation

We quantify slot proximity at the final update using the ratio of the minimum pairwise distance to the mean pairwise distance, and assess whether the closest pair’s predicted segmentation masks overlap.

For each image XnX_n and update step tt, we compute the Euclidean distance between every pair of selected slots:

dn,ij(t)=∥sn,i(t)−sn,j(t)∥2,(i,j)∈Cn,d_{n,ij}^{(t)}=\left\|s_{n,i}^{(t)} - s_{n,j}^{(t)}\right\|_2,\qquad (i,j)\in\mathcal{C}_n,

where Cn\mathcal{C}_n contains all unordered pairs of selected slots in image nn. We then calculate the mean pairwise distance and the minimum pairwise distance for each image:

dˉn(t)=1∣Cn∣∑(i,j)∈Cndn,ij(t),dn,min⁡(t)=min⁡(i,j)∈Cndn,ij(t).\bar{d}_{n}^{(t)} = \frac{1}{\vert \mathcal{C}_n \vert} \sum _{(i,j) \in \mathcal{C}_n} d _{n, ij}^{(t)} , \quad d _{n,\min} ^{(t)} = \min _{(i,j) \in \mathcal{C}_n} d _{n,ij}^{(t)}.

We average the mean and minimum pairwise distances across images as Dpair(T),Dmin(T)D_{\mathrm{pair}}^{(T)} , D_{\mathrm{min}}^{(T)}. At the final update, we report their ratio, Dmin(T)/Dpair(T)D_{\mathrm{min}}^{(T)}/D_{\mathrm{pair}}^{(T)}, to measure the closest-pair distance relative to the overall pairwise distance.

At the final update TT, we identify the pair attaining the minimum distance in each image and compute the IoU between their predicted segmentation masks. This measures whether the closest latent pair also has overlapping masks.

The results show the following:

  1. Slot Attention has a smaller minimum-to-mean distance ratio than SlotODE (0.19 vs. 0.57), indicating that Slot Attention’s closest pair is more compressed, whereas SlotODE’s is more separated, relative to each method’s mean pairwise distance. At the final update, Slot Attention has $D_{\mathrm{pair}}^{(T)}=$5.68 and $D_{\mathrm{min}}^{(T)}=$1.05, whereas SlotODE has $D_{\mathrm{pair}}^{(T)}=$7.68 and $D_{\mathrm{min}}^{(T)}=$4.35.
  2. The closest pair under Slot Attention also has greater overlap between its predicted masks than the closest pair under SlotODE (0.47 vs. 0.10). Together, these results indicate that SlotODE’s closest pair remains more distinct both in latent space and in its predicted masks.
Dmin(T)/Dpair(T)D_{\mathrm{min}}^{(T)} / D_{\mathrm{pair}}^{(T)} (↑\uparrow)IoU of closest pair (↓\downarrow)
Slot Attention0.190.47$_{\pm .02}$
SlotODE0.570.10$_{\pm .01}$

We have newly added Appendix B.5 to the revised manuscript. It summarizes the analysis above and reports Dpair(t)D_{\mathrm{pair}}^{(t)} and Dmin(t)D_{\mathrm{min}}^{(t)} , and their ratio at initialization and every update step.

(2) Slot separation preservation during updates

We also present an additional analysis that measures how the distance between the final closest pair changes from initialization to the final output. At each step, we measure its distance relative to the mean pairwise distance. The following results show stronger preservation of relative slot separation under SlotODE than under Slot Attention.

To measure how the final closest pair’s relative separation changes from initialization to the final update, we compute an image-wise relative separation change ratio RnR_n. For each image XnX_n, we identify the pair with the minimum distance at the final update and track this pair across all update steps. We normalize its distance at each step by the mean pairwise distance among the selected slots in that image:

(in∗,jn∗)=arg⁡min⁡(i,j)∈Cndn,ij(T),rn(t)=dn,in∗jn∗(t)dˉn(t),Rn=rn(T)rn(0).(i_n^\ast, j_n^\ast) =\arg \min_{(i,j)\in\mathcal{C}_n} d _{n,ij}^{(T)} , \qquad r_n^{(t)} = \frac{d _{n,i_n^\ast j_n^\ast}^{(t)}}{\bar d_n^{(t)}} , \qquad R_n=\frac{r_n^{(T)}}{r_n^{(0)}}.

Rn<1R_n<1 indicates that the pair’s distance has decreased relative to the other selected pairs, while Rn>1R_n>1 indicates that its relative distance has increased. We report the mean RnR_n, the fraction of images for which RnR_n falls below each threshold, and the mean IoU of the final masks for those images.

The results show the following:

  1. The mean RnR_n is below 1.0 under Slot Attention (0.65), indicating relative contraction of the final closest pair on average. Under SlotODE, it is above 1.0 (1.82), indicating relative expansion on average.
  2. While 5.41% of Slot Attention images have Rn<0.1R_n<0.1, no SlotODE images fall below this threshold. These results indicate that the final closest pair retains stronger relative separation under SlotODE.
  3. Among images with Rn<0.7R_n<0.7, the final closest pair’s mean mask IoU is lower under SlotODE than under Slot Attention (0.23 vs. 0.32). Slot Attention images with the more severe Rn<0.1R_n<0.1 have a mean IoU of 0.86, whereas SlotODE’s rare Rn<0.7R_n<0.7 cases have a mean IoU of 0.23. Stronger preservation of relative separation under SlotODE is accompanied by lower mask overlap.
MethodMean Rn(↑)R_n (\uparrow)
Slot Attention0.65
SlotODE1.82
MethodRn<0.1R_n<0.1 (%)(↓)(\downarrow)IoU (Rn<0.1)(↓)(R_n<0.1)(\downarrow)Rn<0.7R_n<0.7 (%)(↓)(\downarrow)IoU (Rn<0.7)(↓)(R_n<0.7) (\downarrow)
Slot Attention5.41%0.86$_{\pm .10}$56.99%0.32$_{\pm .16}$
SlotODE0%--0.58%0.23$_{\pm .15}$

We have newly added Appendix B.6 to the revised manuscript. It reports results at each update step and visualizes the image-wise distribution of RnR_n. We would appreciate the reviewer’s consideration of these additional results.

W1-2 Numerical agreement between three-step Euler and reference solutions

We agree that the continuous-time guarantee does not automatically extend to the implemented three-step Euler update. In response to a related question from reviewer 9sTg, we compare the implemented three-step Euler update with high-accuracy numerical references obtained using Dopri5 and RK4 on MOVi-C. All solvers use the same trained vector field and initial slot states. We compare their outputs at matched times t=0,1/3,2/3,1t=0,1/3,2/3,1. Euler always uses three steps; RK4 uses either 3 or 12 integration steps. For Dopri5, 3 and 12 denote uniform output intervals; its internal steps are adaptive. The ∗^* marks Dopri5 outputs evaluated on a uniform grid.

Our analysis shows that:

  • At the final update, Euler’s slot states closely match the reference states (cosine similarity 0.9993; relative ℓ2\ell_2 error approximately 0.032).
  • Euler also closely matches the reference pairwise slot differences, with relative error of approximately 0.02 across all pairs and 0.06 for the closest pairs.
  • Each slot’s final mask from Euler overlaps substantially with its corresponding mask from the reference solver (mean IoU 0.86–0.87).

(1) Slot-state agreement

We first test whether Euler reaches slot states close to the reference solutions. At matched times we measure (1) slot-state cosine similarity and (2) relative ℓ2\ell_2 error, defined as ∥sEuler−sref∥2∥sEuler∥2\frac{\Vert s_{\mathrm{Euler}}-s_{\mathrm{ref}} \Vert_2}{\Vert s_{\mathrm{Euler}} \Vert_2}. The table reports results at the final time t=1t=1.

The results show the following:

  1. At t=1t=1, the three-step Euler states closely agree with both reference solutions, with cosine similarity of 0.9993 and relative ℓ2\ell_2 error of about 0.032. Similar agreement is observed with 12 RK4 steps and 12 Dopri5 output intervals. Thus, the implemented Euler update closely approximates the reference slot states at the final time.
Reference solverReference timeCosine similarity (↑\uparrow)Relative ℓ2\ell_2 error (↓\downarrow)
Dopri53/3$^*$0.9993$_{\pm .00}$0.0318$_{\pm .02}$
RK43/30.9993$_{\pm .00}$0.0315$_{\pm .02}$
Dopri512/12$^*$0.9993$_{\pm .00}$0.0318$_{\pm .02}$
RK412/120.9993$_{\pm .00}$0.0318$_{\pm .02}$

(2) Pairwise slot separation agreement

We next test whether Euler reproduces the differences between slot states, particularly for pairs that are closest under Euler. For each candidate slot pair (i,j)(i,j) and solver qq, let Δn,ijq=sn,iq−sn,jq\Delta_{n,ij}^{q}=s_{n,i}^{q}-s_{n,j}^{q}. We measure the relative ℓ2\ell_2 error ∥Δn,ijEuler−Δn,ijref∥2∥Δn,ijEuler∥2\frac{\Vert \Delta_{n,ij}^{\mathrm{Euler}} - \Delta_{n,ij}^{\mathrm{ref}} \Vert_2}{ \Vert \Delta_{n,ij}^{\mathrm{Euler}} \Vert_2} . We report this error for all candidate pairs (All), the closest pair under Euler (Top-1), and the five closest pairs under Euler (Top-5). Pair identities are held fixed across solvers.

The results show the following:

  1. At t=1t=1, the relative error is about 0.02 across all pairs and about 0.06 for the closest pairs under both reference solvers. The three-step Euler update therefore approximates the reference pairwise slot differences on average, including those of the closest pairs.
Reference solverReference timeAll (↓\downarrow)Top-1 (↓\downarrow)Top-5 (↓\downarrow)
Dopri53/3$^*$0.0238$_{\pm .01}$0.0623$_{\pm .08}$0.0572$_{\pm .04}$
RK43/30.0233$_{\pm .01}$0.0615$_{\pm .08}$0.0562$_{\pm .04}$
Dopri512/12$^*$0.0238$_{\pm .01}$0.0623$_{\pm .08}$0.0572$_{\pm .04}$
RK412/120.0238$_{\pm .01}$0.0623$_{\pm .08}$0.0572$_{\pm .04}$

(3) Final mask agreement

Finally, we test whether the numerical differences lead to different decoded segmentation masks. We compare each slot’s final mask from Euler with the corresponding mask from each reference solution. The table reports mask IoU as the mean and standard deviation across images and slots; all rows correspond to t=1t=1.

The results show the following:

  1. The final masks from three-step Euler have a mean IoU of approximately 0.86-0.87 with those from both reference solvers. The values remain similar with 12 reference intervals. The agreement observed in slot states and pairwise differences is therefore also reflected in the predicted masks on average.
Reference solverReference timeIoU (↑\uparrow)
Dopri53/3$^*$0.8691$_{\pm .19}$
RK43/30.8654$_{\pm .20}$
Dopri512/12$^*$0.8681$_{\pm .19}$
RK412/120.8632$_{\pm .20}$

We have added Appendix B.7 to the revised manuscript, where we summarize this solver-agreement analysis and report additional results at intermediate update times. We would appreciate the reviewer’s consideration of these results.

Reviewer_9sTgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies slots that become identical during iterative updates. The authors propose SlotODE. It models the joint slot state with a permutation-equivariant ODE and learns its time derivative. Under the stated assumptions, the paper proves that the exact ODE solution has no finite-time slot-state collision. It also gives an exponential lower bound on pairwise slot separation. Experiments on four datasets show gains over conventional Slot Attention.

Critical strengths
  • The idea is clear. It combines the uniqueness of ODE solutions with permutation equivariance. This gives an explicit theoretical explanation of slot-state separation. The proofs are valid under the stated assumptions.
  • The paper analyzes the update form, the initialization, and spectral normalization. Performance remains similar with a different solver or more test-time integration steps.
Critical weaknesses

W1. The theoretical guarantee has not yet been carried over to the actual numerical update. The theorems concern the exact ODE solution. The actual model uses a three-step Euler update. The existing solver experiments mainly report segmentation performance. But it does not directly test whether the actual update keeps slots separated.

W2. There's no direct evidence linking reduced collisions to performance gains. Figures 1–2 show 2D PCA projections. Overlap in the projection does not prove that the original slot states are identical. Distinct slot states can also represent the same object. The paper’s ablations support the effectiveness of the proposed design. However, they do not yet link collision reduction to better segmentation.

Questions for authors

Q1. Can you report pairwise slot distances and the fraction of near-overlapping slots in the original representation space? Can you also check how these relate to duplicate masks and segmentation errors?

Q2. Can you compare the three-step Euler update with a high-accuracy reference solution, in terms of state error and degree of separation? This would help judge the link between the continuous theory and the actual implementation.

Q3. Please specify the actual selection loss, normalization scheme and coefficient for SA* and SlotODE. Please also make the results in the main text and the appendix consistent. If the settings differ, please add a comparison under the same setting.

Q4. Some of the experimental set-ups and results need to be made consistent. Appendix D.2 sets the regularisation coefficient of the AdaSlot baseline at 0.5. D.1 and Table 16 however set the SlotODE coefficient at 1.0 on the complex datasets. It's not clear whether they are using the same selection loss and weight. There's also an inconsistency with the FG-ARI of SPOT-MLP on COCO: it's 41.95 in Table 2 and 40.35 in Table 22. That impacts reproducibility and the interpretation of the results.

5 comments

We are grateful to Reviewer 9sTg for the thoughtful comments and valuable feedback on our manuscript.

W1 & W2 Slot separation and mask overlap during Euler updates

To address both concerns, we compare Slot Attention and SlotODE on MOVi-C under the same adaptive slot-selection mechanism. Our two numerical analyses show that SlotODE better preserves the relative separation of the final closest slot pair and exhibits less mask overlap than Slot Attention, consistent with its better segmentation performance.

(1) Quantitative analysis of slot separation

We quantify slot proximity at the final update using the ratio of the minimum pairwise distance to the mean pairwise distance, and assess whether the closest pair’s predicted segmentation masks overlap.

For each image XnX_n and update step tt, we compute the Euclidean distance between every pair of selected slots:

dn,ij(t)=∥sn,i(t)−sn,j(t)∥2,(i,j)∈Cn,d_{n,ij}^{(t)}=\left\|s_{n,i}^{(t)} - s_{n,j}^{(t)}\right\|_2,\qquad (i,j)\in\mathcal{C}_n,

where Cn\mathcal{C}_n contains all unordered pairs of selected slots in image nn. We then calculate the mean pairwise distance and the minimum pairwise distance for each image:

dˉn(t)=1∣Cn∣∑(i,j)∈Cndn,ij(t),dn,min⁡(t)=min⁡(i,j)∈Cndn,ij(t).\bar{d}_{n}^{(t)} = \frac{1}{\vert \mathcal{C}_n \vert} \sum _{(i,j) \in \mathcal{C}_n} d _{n, ij}^{(t)} , \quad d _{n,\min} ^{(t)} = \min _{(i,j) \in \mathcal{C}_n} d _{n,ij}^{(t)}.

We average the mean and minimum pairwise distances across images as Dpair(T),Dmin(T)D_{\mathrm{pair}}^{(T)} , D_{\mathrm{min}}^{(T)}. At the final update, we report their ratio, Dmin(T)/Dpair(T)D_{\mathrm{min}}^{(T)}/D_{\mathrm{pair}}^{(T)}, to measure the closest-pair distance relative to the overall pairwise distance.

At the final update TT, we identify the pair attaining the minimum distance in each image and compute the IoU between their predicted segmentation masks. This measures whether the closest latent pair also has overlapping masks.

The results show the following:

  1. Slot Attention has a smaller minimum-to-mean distance ratio than SlotODE (0.19 vs. 0.57), indicating that Slot Attention’s closest pair is more compressed, whereas SlotODE’s is more separated, relative to each method’s mean pairwise distance. At the final update, Slot Attention has $D_{\mathrm{pair}}^{(T)}=$5.68 and $D_{\mathrm{min}}^{(T)}=$1.05, whereas SlotODE has $D_{\mathrm{pair}}^{(T)}=$7.68 and $D_{\mathrm{min}}^{(T)}=$4.35.
  2. The closest pair under Slot Attention also has greater overlap between its predicted masks than the closest pair under SlotODE (0.47 vs. 0.10). Together, these results indicate that SlotODE’s closest pair remains more distinct both in latent space and in its predicted masks.
Dmin(T)/Dpair(T)D_{\mathrm{min}}^{(T)} / D_{\mathrm{pair}}^{(T)} (↑\uparrow)IoU of closest pair (↓\downarrow)
Slot Attention0.190.47$_{\pm .02}$
SlotODE0.570.10$_{\pm .01}$

We have newly added Appendix B.5 to the revised manuscript. It summarizes the analysis above and reports Dpair(t)D_{\mathrm{pair}}^{(t)} and Dmin(t)D_{\mathrm{min}}^{(t)} , and their ratio at initialization and every update step.

(2) Slot separation preservation during updates

We also present an additional analysis that measures how the distance between the final closest pair changes from initialization to the final output. At each step, we measure its distance relative to the mean pairwise distance. The following results show stronger preservation of relative slot separation under SlotODE than under Slot Attention.

To measure how the final closest pair’s relative separation changes from initialization to the final update, we compute an image-wise relative separation change ratio RnR_n. For each image XnX_n, we identify the pair with the minimum distance at the final update and track this pair across all update steps. We normalize its distance at each step by the mean pairwise distance among the selected slots in that image:

(in∗,jn∗)=arg⁡min⁡(i,j)∈Cndn,ij(T),rn(t)=dn,in∗jn∗(t)dˉn(t),Rn=rn(T)rn(0).(i_n^\ast, j_n^\ast) =\arg \min_{(i,j)\in\mathcal{C}_n} d _{n,ij}^{(T)} , \qquad r_n^{(t)} = \frac{d _{n,i_n^\ast j_n^\ast}^{(t)}}{\bar d_n^{(t)}} , \qquad R_n=\frac{r_n^{(T)}}{r_n^{(0)}}.

Rn<1R_n<1 indicates that the pair’s distance has decreased relative to the other selected pairs, while Rn>1R_n>1 indicates that its relative distance has increased. We report the mean RnR_n, the fraction of images for which RnR_n falls below each threshold, and the mean IoU of the final masks for those images.

The results show the following:

  1. The mean RnR_n is below 1.0 under Slot Attention (0.65), indicating relative contraction of the final closest pair on average. Under SlotODE, it is above 1.0 (1.82), indicating relative expansion on average.
  2. While 5.41% of Slot Attention images have Rn<0.1R_n<0.1, no SlotODE images fall below this threshold. These results indicate that the final closest pair retains stronger relative separation under SlotODE.
  3. Among images with Rn<0.7R_n<0.7, the final closest pair’s mean mask IoU is lower under SlotODE than under Slot Attention (0.23 vs. 0.32). Slot Attention images with the more severe Rn<0.1R_n<0.1 have a mean IoU of 0.86, whereas SlotODE’s rare Rn<0.7R_n<0.7 cases have a mean IoU of 0.23. Stronger preservation of relative separation under SlotODE is accompanied by lower mask overlap.
MethodMean Rn(↑)R_n (\uparrow)
Slot Attention0.65
SlotODE1.82
MethodRn<0.1R_n<0.1 (%)(↓)(\downarrow)IoU (Rn<0.1)(↓)(R_n<0.1)(\downarrow)Rn<0.7R_n<0.7 (%)(↓)(\downarrow)IoU (Rn<0.7)(↓)(R_n<0.7) (\downarrow)
Slot Attention5.41%0.86$_{\pm .10}$56.99%0.32$_{\pm .16}$
SlotODE0%--0.58%0.23$_{\pm .15}$

We have newly added Appendix B.6 to the revised manuscript. It reports results at each update step and visualizes the image-wise distribution of RnR_n. We would appreciate the reviewer’s consideration of these additional results.

Q1 Evidence connecting slot separation and mask overlap

We would appreciate it if the reviewer could consider our responses to W1 & W2, which show that:

  • At the final update, SlotODE’s closest pair is more separated relative to the mean pairwise distance than Slot Attention’s (0.57 vs. 0.19), and its predicted masks overlap less (IoU 0.10 vs. 0.47).
  • From initialization to the final update, this pair’s relative separation increases on average under SlotODE but decreases under Slot Attention (mean RnR_n: 1.82 vs. 0.65).

Q2 Agreement between three-step Euler and reference solvers

We compare the implemented three-step Euler update with high-accuracy numerical references obtained using Dopri5 and RK4 on MOVi-C. All solvers use the same trained vector field and initial slot states. We compare their outputs at matched times t=0,1/3,2/3,1t=0,1/3,2/3,1. Euler always uses three steps; RK4 uses either 3 or 12 integration steps. For Dopri5, 3 and 12 denote uniform output intervals; its internal steps are adaptive. The ∗^* marks Dopri5 outputs evaluated on a uniform grid.

Our analysis shows that:

  • At the final update, Euler’s slot states closely match the reference states (cosine similarity 0.9993; relative ℓ2\ell_2 error approximately 0.032).
  • Euler also closely matches the reference pairwise slot differences, with relative error of approximately 0.02 across all pairs and 0.06 for the closest pairs.
  • Each slot’s final mask from Euler overlaps substantially with its corresponding mask from the reference solver (mean IoU 0.86–0.87).

(1) Slot-state agreement

We first test whether Euler reaches slot states close to the reference solutions. At matched times we measure (1) slot-state cosine similarity and (2) relative ℓ2\ell_2 error, defined as ∥sEuler−sref∥2∥sEuler∥2\frac{\Vert s_{\mathrm{Euler}}-s_{\mathrm{ref}} \Vert_2}{\Vert s_{\mathrm{Euler}} \Vert_2}. The table reports results at the final time t=1t=1.

The results show the following:

  1. At t=1t=1, the three-step Euler states closely agree with both reference solutions, with cosine similarity of 0.9993 and relative ℓ2\ell_2 error of about 0.032. Similar agreement is observed with 12 RK4 steps and 12 Dopri5 output intervals. Thus, the implemented Euler update closely approximates the reference slot states at the final time.
Reference solverReference timeCosine similarity (↑\uparrow)Relative ℓ2\ell_2 error (↓\downarrow)
Dopri53/3$^*$0.9993$_{\pm .00}$0.0318$_{\pm .02}$
RK43/30.9993$_{\pm .00}$0.0315$_{\pm .02}$
Dopri512/12$^*$0.9993$_{\pm .00}$0.0318$_{\pm .02}$
RK412/120.9993$_{\pm .00}$0.0318$_{\pm .02}$

(2) Pairwise slot separation agreement

We next test whether Euler reproduces the differences between slot states, particularly for pairs that are closest under Euler. For each candidate slot pair (i,j)(i,j) and solver qq, let Δn,ijq=sn,iq−sn,jq\Delta_{n,ij}^{q}=s_{n,i}^{q}-s_{n,j}^{q}. We measure the relative ℓ2\ell_2 error ∥Δn,ijEuler−Δn,ijref∥2∥Δn,ijEuler∥2\frac{\Vert \Delta_{n,ij}^{\mathrm{Euler}} - \Delta_{n,ij}^{\mathrm{ref}} \Vert_2}{ \Vert \Delta_{n,ij}^{\mathrm{Euler}} \Vert_2} . We report this error for all candidate pairs (All), the closest pair under Euler (Top-1), and the five closest pairs under Euler (Top-5). Pair identities are held fixed across solvers.

The results show the following:

  1. At t=1t=1, the relative error is about 0.02 across all pairs and about 0.06 for the closest pairs under both reference solvers. The three-step Euler update therefore approximates the reference pairwise slot differences on average, including those of the closest pairs.
Reference solverReference timeAll (↓\downarrow)Top-1 (↓\downarrow)Top-5 (↓\downarrow)
Dopri53/3$^*$0.0238$_{\pm .01}$0.0623$_{\pm .08}$0.0572$_{\pm .04}$
RK43/30.0233$_{\pm .01}$0.0615$_{\pm .08}$0.0562$_{\pm .04}$
Dopri512/12$^*$0.0238$_{\pm .01}$0.0623$_{\pm .08}$0.0572$_{\pm .04}$
RK412/120.0238$_{\pm .01}$0.0623$_{\pm .08}$0.0572$_{\pm .04}$

(3) Final mask agreement

Finally, we test whether the numerical differences lead to different decoded segmentation masks. We compare each slot’s final mask from Euler with the corresponding mask from each reference solution. The table reports mask IoU as the mean and standard deviation across images and slots; all rows correspond to t=1t=1.

The results show the following:

  1. The final masks from three-step Euler have a mean IoU of approximately 0.86-0.87 with those from both reference solvers. The values remain similar with 12 reference intervals. The agreement observed in slot states and pairwise differences is therefore also reflected in the predicted masks on average.
Reference solverReference timeIoU (↑\uparrow)
Dopri53/3$^*$0.8691$_{\pm .19}$
RK43/30.8654$_{\pm .20}$
Dopri512/12$^*$0.8681$_{\pm .19}$
RK412/120.8632$_{\pm .20}$

We have added Appendix B.7 to the revised manuscript, where we summarize this solver-agreement analysis and report additional results at intermediate update times. We would appreciate the reviewer’s consideration of these results.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.