Bootstrapping Top-down Information for Self-modulating Slot Attention

Object-centric learning (OCL) aims to learn representations of individual objects within visual scenes without manual supervision, facilitating efficient and effective visual reasoning. Traditional OCL methods primarily employ bottom-up approaches that aggregate homogeneous visual features to represent objects. However, in complex visual environments, these methods often fall short due to the heterogeneous nature of visual features within an object. To address this, we propose a novel OCL framework incorporating a top-down pathway. This pathway first bootstraps the semantics of individual objects and then modulates the model to prioritize features relevant to these semantics. By dynamically modulating the model based on its own output, our top-down pathway enhances the representational quality of objects. Our framework achieves state-of-the-art performance across multiple synthetic and real-world object-discovery benchmarks.

Paper

Similar papers

Peer review

Reviewer PJTV5/10 · confidence 4/52024-07-07

Summary

This paper proposes a method that improves the performance of object-centric learning by modulating Slot Attention with semantic and location information obtained based on the output slots and attention maps of Slot Attention. For a given output slot, the semantic information is chosen as the vector closest to this slot in a codebook that is learned from all output slots of the entire dataset via vector quantization, and the location information is chosen as the shifted attention map with a mean value of 1. The proposed method is implemented based on DINASOUR and is compared with DINASOUR and other methods on two synthetic and two real datasets. The proposed method outperforms the compared method in most cases when the size of codebook is chosen appropriately.

Strengths

1. Improving the performance of object-centric learning with top-down information is an important and interesting research direction. 2. The proposed method outperforms the compared methods when the size of codebook is chosen appropriately.

Weaknesses

1. My main concern about the paper is that the performance of the proposed method is very sensitive to the size of codebook. If the size of codebook is not chosen appropriately, the performance can be even worse than the original DINASOUR method. Moreover, how to choose the best size of codebook is not described clearly. 2. From my understanding, the core of the proposed method is a self-modulation module (along with a codebook) that is compatible with all the object-centric learning methods using the slot attention mechanism. However, the proposed method is only implemented based on DINASOUR. The quality of the paper could be significantly improved if the proposed method is implemented based on at least another object-centric learning method that is developed based on slot attention. 3. The proposed self-modulation includes both semantic and spatial modulations. However, the codebook used in the semantic modulation also contains spatial information. This design can be more elegant, and the codebook can be much more useful if it only contains semantic information. In this way, whether two objects belong to the same category can be determined automatically. 4. Some data in the experimental part are inconsistent. For example, the mBO^i of the proposed method on COCO is 33.0 in Table 2, 33.3 in Table 3, and 32.7 in Table 4. 5. The authors attribute both semantic modulation and spatial modulation as top-down modulations. I agree that the semantic modulation is a top-down modulation because the codebook contains knowledge learned from all images (not just the inferred image). However, I don’t think that the spatial modulation can be considered top-down because it just uses attention maps computed in a bottom-up manner (only based on the content of one image).

Questions

1. How is the best size of codebook chosen? 2. Is it possible to learn a codebook that is independent of the locations of objects in the image?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have addressed the limitation in the conclusion part, and there is no potential negative societal impact.

Reviewer qEia6/10 · confidence 4/52024-07-11

Summary

Slot Attention is a popular component for object-centric learning methods. In this paper, the authors propose an extension of Slot Attention named "top-down pathway". After the last iteration of Slot Attention, the slots are mapped to a discrete, learnable codebook. Jointly with the final attention map, this information is used in a second iteration of Slot Attention to modulate the QKV-like attention mechanism. Previous models are consistently outperformed on established benchmarks. Moreover, a more detailed analysis of the learned codebook shows that distinct semantic concepts are represented by different codebook entries.

Strengths

- The proposed extension of Slot Attention is clearly explained. The original Slot Attention is described at an adequate level of detail, which helps to make the paper accessible. - The model consistently outperforms DINOSAUR, from which all design decisions are derived. This demonstrates the effectiveness of the novel mechanism. - The analysis of the learned codebook is insightful and confirms the motivation of the proposed method.

Weaknesses

I am not fully convinced by the term "top-down pathway". Only the recurrent Slot Attention module has been extended, the rest of the model is still bottom-up. From the initial description, I would have expected top-down modulation reaching to the encoder. Did you consider a variant of the model that continues with the slot embeddings from the first iteration instead of starting from scratch? If this performs similarly well, it might indiciate that the quantization due to the codebook is driving the improvement.

Questions

- A recent approach to object-centric learning which is not based on Slot Attention is CutLER (Wang et al. 2023). How does the proposed method compare to CutLER?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations are very briefly discussed. An additional point could be the dependence on the pretrained feature encoder, which might not work well in all scenarios.

Reviewer Yith7/10 · confidence 3/52024-07-12

Summary

This paper proposes a modification to Slot Attention incorporating top-down information into the algorithm. After an iteration of Slot Attention, the slots are quantized into a learned codebook. The quantized slots and attention maps are then used in another iteration of Slot Attention, refining the representations. The algorithm is evaluated in the DINOSAUR setting on the MOVi, Pascal VOC, and COCO datasets, showing improved segmentation quality over vanilla DINOSAUR. Visualizations of the codebook show that meaningful semantic concepts are learned and several ablations are performed.

Strengths

The proposed method is well-motivated and seems to show an improvement over previous methods. The use of top-down semantic information from a learned codebook is novel, from my understanding. Overall, the paper is well-written and the authors provide a good analysis of their method including ablations of the different design choices.

Weaknesses

1. The experiments are only performed in the DINOSAUR setting with pretrained, frozen ViT features as the input and reconstruction target. It would be very informative to also include experiments with the original image as the input and reconstruction target since that is also a common use case for Slot Attention. By only evaluating in the DINOSAUR setting, it is unclear how reliant the performance of the proposed model is on the pretrained ViT features. 2. In Figure 3 and 5.B, it seems the slots are sometimes reassigned before and after the self-modulation step? For example, the skater in Figure 3 is captured by the 4th slot before the modulation and the 5th slot after the modulation. The 5th slot before the modulation seems to capture part of the background, not any part of the skater. It is unclear to me why this would happen during the modulation update.

Questions

1. Since the slots also contain position information (”where” information), it seems possible that this gets captured in the learned codebook. Is this something the authors observed? 2. How cherry-picked is Figure 2? Are there cases where these codes do not correspond to the same semantic concept? 3. In Figure 2, it seems that sometimes, multiple objects with similar semantics (e.g. multiple zebras or signs) are being captured by one slot. Did you notice your proposed method does more semantic grouping instead of instance grouping compared with vanilla DINOSAUR?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer 3ckb6/10 · confidence 4/52024-07-12

Summary

This work proposes a novel unsupervised object-centric learning method. In particular, the proposed approach enhances the slot-attention mechanism by incorporating a "top-down pathway" that highlights features relevant to objects in the image. The method first employs a standard self-attention mechanism to extract slot vectors and the corresponding attention masks. These initial slot vectors and attention masks are then fed into another round of the slot-attention mechanism. In this round, they modulate the value maps of the cross-attention mechanism separately for each slot. This modulation is channel-wise for slot vectors and spatial-wise for attention masks, enabling the slot-attention mechanism to emphasize to features relevant to the slots discovered on the first slot-attention round. Furthermore, the method quantizes the slot vectors from the first slot-attention round using the VQ approach before utilizing them for modulation in the second round. To evaluate the proposed self-modulated-based slot-attention mechanism, the authors integrate it into the DINOSAUR framework and apply it to the COCO, VOC, MOVIE-E, and MOVIE-C datasets for the downstream task of object segmentation. The results demonstrate improvements over the vanilla slot-attention mechanism. Ablation studies are conducted on the COCO dataset.

Strengths

STRENGTHS: - The idea of modulating the value maps for each slot, using slots found from a previous self-attention round, is interesting. This technique allows the subsequent self-attention round to focus more on features that are relevant to the objects discovered in the initial round. Essentially, it's an additional iterative improvement for the extracted slots, supplementing the existing iterative mechanism in the standard slot-attention. - The proposed method has been demonstrated to improve object segmentation results on all tested datasets. Furthermore, the authors utilized challenging datasets, such as COCO, to validate the effectiveness of their method, rather than resorting to simpler one, which is often among other object-centric works. - The paper has strong results and detailed ablation studies.

Weaknesses

- **(W1)** It's unusual that the mIoU results in the MOVIE-C and MOVIE-E datasets surpass the mBO results. The mBO is computed by assigning each ground truth mask the predicted mask with the largest overlap, and then averaging the IoUs of the assigned mask pairs. On the other hand, the mIoU metric is more strict and employs Hungarian matching (rather than greedy matching) to assign predicted masks to the ground truth masks. Consequently, mBO results should be either greater than or equal to the mIoU results. Therefore, it's highly likely that there's a bug in the computation of the mBO or the mIoU metrics on the MOVIE-C and MOVIE-E datasets. Or a typo / mistake when copying the results to the paper maybe. - As a relevant side note for the upcoming comments, it's important to mention that the FG-ARI metric is generally viewed as unreliable for assessing unsupervised object-centric methods. This is because it only takes foreground pixels into account, which can provide a deceptive understanding of segmentation quality by disregarding the localization accuracy of predicted masks [A, B, C, 30, 42]. - **(W2)** The motivation behind the Vector Quantization (VQ) component of the method is unclear. In particular it's uncertain why quantizing the slot vectors is essential for self-modulation. In fact, Table 4 indicates that the method is sensitive to the codebook size (also acknowledged by the authors in the limitations section). The mBO performance for sub-optimal codebook sizes 128, 256, and 1024 (30.2%, 30.5%, and 29.8%, respectively) is either the same or worse than the reproduced DINOSAUR mBO performance (30.5% from Table 6). Conversely, removing VQ achieves 32.3%, which is only marginally worse than the 32.7% obtained with VQ and the optimal codebook size. Therefore, VQ unnecessarily complicates the proposed method. As mentioned earlier, FG-ARI is unreliable, and the higher differences in this metric (36.3% w/o VQ vs 37.7% w/ VQ) should be disregarded. - It's also somewhat concerning that Table 6 shows only a minimal impact on mBO performance when using channel-wise modulation (from 32.5% to 32.7%). However, this does not really increase the complexity of the method, as VQ does. - **(W3)** Lacks discussions on how the proposed method affects the training and test time - **(W4)** Table 3 is missing comparisons with related works, such as Rotating Features [D] and SPOT [C], which also demonstrate strong results. [A] Genesis: Generative scene inference and sampling with object-centric latent representations, ICLR 2020. [B] Unsupervised Layered Image Decomposition into Object Prototypes, ICCV 2021. [C] SPOT: Self-Training with Patch-Order Permutation for Object-Centric Learning with Autoregressive Transformers, CVPR 2024 [D] Rotating features for object discovery. NeurIPs 2023. [30] Bridging the gap to real-world object-centric learning, ICLR 2023. [42] Slotdiffusion: Object-centric generative modeling with diffusion models, NeurIPs 2023.

Questions

- My primary concern regarding this work (mentioned as (W2) in the weaknesses section) is that the channel-wise and Vector Quantization (VQ) components unnecessarily complicate the method without demonstrating significant improvements in the ablation results. I would appreciate it if the authors could provide compelling arguments, such as additional experimental evidence on other datasets (e.g., VOC or MOVIE-E/C) or different settings (e.g., longer training), that show (if it is the case) the benefits of channel-wise modulation with VQ. - Please address the concerns (W1), (W3), and (W4) mentioned in the weaknesses section.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer PJTV2024-08-10

Thanks for the detailed response. Since the codebook size is determined solely based on the training set in an unsupervised way and applying the proposed self-modulation to Slot Attention also improves the performance on the CLEVR6 dataset, my concerns about technical flaws disappear and the rating is increased.

Authorsrebuttal2024-08-13

We sincerely appreciate your thoughtful reconsideration of our work. We are pleased that our clarifications regarding codebook size determination and the CLEVR6 experiment have addressed your concerns. We will improve the paper in future revision based on your feedback. Thank you for your valuable feedback and updated assessment.

Reviewer qEia2024-08-12

Thank you very much for the detailed response. - I am still not entirely convinced about the tmerin "top-down pathway", however I think this not a major issue that speaks against accepting the paper. - The additional comparisons to MaskCut and other methods in the responses to other reviewers are very helpful and confirm the consistent improvements of the proposed model. - I agree with the other reviewers on the concerns regarding the codebook size and the unclear role of the vector quantization. (1) While the codebook size can be chosen without supervision, it seems that the required multiple training runs substantially increase the computational cost. (2) The impact of the codebook on the performance seems to be small and heavily decreases the performance when used alone, as shown by the additional ablation study. While I still think the analysis of the codebook is interesting, I am not sure to which extent it is necessary for the performance improvement. Overall, I still think the paper should be accepted, since the proposed method is interesting and consistently outperforms previous methods. Due to the remaining concerns I keep my rating for now. But I am looking forward to read the other reviewers comments on these points and happy to discuss further.

Authorsrebuttal2024-08-13

We sincerely appreciate your detailed response and thoughtful consideration of our rebuttal. We will address your feedback as follows: - We will elaborate on the term "top-down pathway" in the paper to clarify its usage and context. - We're pleased you found the additional comparisons to MaskCut helpful. We will include this in the future revision. - Regarding codebook concerns: a) While determining optimal codebook size requires multiple training runs, perplexity is measured at half the full training schedule (250K), reducing computational burden. b) On the necessity of vector quantization, we kindly refer you to our response to reviewer 3ckb, specifically points (1), (3), and (4) under **"Necessity & effect of VQ"**, which provide further insights into VQ's role in our method's performance. We truly appreciate your valuable feedback and will incorporate your suggestions to improve our paper's final version. Thank you for your support!

Reviewer Yith2024-08-12

Reply to Rebuttal

Thank you for the rebuttal and running additional experiments. The updated figures for the slots before and after modulation make much more sense now. The additional results with vanilla Slot Attention on CLEVR are also promising, although I would encourage the authors to run additional seeds and/or make sure the hyperparameters are correct, as I've seen better results with vanilla slot attention on CLEVR (looking at the qualitative results it seems the background is being split). For what it's worth, I think it would still be a useful contribution if the proposed approach only improves the metrics on more complex datasets (since as a field, we should be generally pushing towards more complex datasets), but it is important to know the effect on simpler datasets as well so researchers are aware of the limitations. I would encourage the authors to include these results in the final version of the paper. After reading the rebuttal as well as the other reviews and rebuttals, I have decided to increase my score to 7.

Authorsrebuttal2024-08-13

We appreciate your thorough review and time evaluating our rebuttal. We're pleased our updated figures and CLEVR experiments have resolved your concerns. We acknowledge your suggestion and will include CLEVR results from multiple seeds in the final paper. Thank you for your constructive feedback and recognizing the strength of our work.

Reviewer 3ckb2024-08-13

I thank the authors for the detailed responses to my comments. Their response addresses my main concerns regarding the necessity of VQ: the codebook size selection is automatic and uses the training split, and VQ offers bigger improvement for longer training. Therefore, and after reading the other reviews and the respective rebuttals, I am going to increase my score.

Authorsrebuttal2024-08-14

We sincerely appreciate your careful review of our responses and your reconsideration. We're glad our explanations about codebook size selection and VQ benefits clarified your concerns. In our future revision, we will include discussions and experiments about VQ, which will significantly improve the paper's quality. Thank you for your valuable feedback!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC