Pith. sign in

REVIEW 3 major objections 4 minor 26 references

ORGAN, a CycleGAN that translates images into object lists and back, is the only tested approach that reliably detects separate objects in dense, low-contrast real-world scenes such as microscopy images.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 19:27 UTC pith:L7STVWJN

load-bearing objection Clever CycleGAN-based object-centric model with a real architectural contribution, but the paper's unsupervised claim rests on an unspecified target in Eq. (3) that may leak object locations. the 3 major comments →

arxiv 2603.02063 v2 pith:L7STVWJN submitted 2026-03-02 cs.CV

ORGAN: Object-Centric Representation Learning using Cycle Consistent Generative Adversarial Networks

Jo\"el K\"uchler , Ellen van Maren , Vaiva Vasiliauskait\.e , Katarina Vuli\'c , Reza Abbasi-Asl , Stephan J. Ihle This is my paper
classification cs.CV MSC 68T4568T07
keywords object-centric representation learningCycleGANunsupervised object detectionimage-to-list translationmicroscopy image segmentationdifferentiable top-kcycle consistencylow-contrast scenes
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes ORGAN, an object-centric representation learner that replaces the usual autoencoder backbone with a cycle-consistent GAN translating between two domains: raw images and structured lists of objects. The authors aim to show that this image-to-list translation can detect and represent individual objects in scenes where existing slot-attention and spatial-attention models fail—specifically images with many small, visually similar, low-contrast objects such as microscopy cells. They report that ORGAN matches prior models on synthetic benchmarks, is the only tested method that detects separate objects in the cell dataset, and that a model trained on 128×128 crops can be applied to 768×768 images without retraining. If true, this would give GAN-based cycle-consistency a practical edge in real-world object-centric learning.

Core claim

ORGAN's central claim is that an unpaired image-to-list translation objective, enforced by two GANs and a cycle-consistency loss, yields object-centric representations that are competitive on synthetic data and superior on challenging real-world low-contrast multi-object scenes. The model detects objects by partitioning the image into overlapping patches, scoring them, suppressing non-maxima, and selecting the top k patches with a differentiable operator; each selected patch is encoded into a list entry with position, presence probability, and feature vector. The reverse generator places Gaussian blobs from the list into a multi-channel image and refines it with a convolutional network. The

What carries the argument

The key mechanism is the cycle-consistent generative adversarial network with two domains, image and list. The image-to-list generator uses a differentiable top-k patch selection after non-maximum suppression, so the model can output a variable number of object proposals; the list-to-image generator turns each list entry into an isotropic Gaussian blob whose channel values are projections of the feature vector, then a U-Net-style convolutional network renders the final image. Training is driven by adversarial losses for each domain plus two cycle-consistency terms—mean absolute error for images and a matched element-wise loss for lists, where matching is solved by a linear-sum assignment (Hu

Load-bearing premise

The training signal includes a term that requires the object locations to compute a target score map; if these locations come from ground-truth labels, the method is not fully unsupervised, and if they are derived from the model's own output, the cycle-consistency signal becomes circular—the paper does not specify which.

What would settle it

Run ORGAN on a dataset with no available object locations and a fixed, known object count; if the model cannot be trained without the t_p term, or if detection fails when t_p is removed, the claimed unsupervised capability is not established. Alternatively, compute t_p from the model's own patch scores and show that the cycle loss is minimized even when object detection is wrong.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If ORGAN's scaling claim holds, a single model trained on small patches can be deployed on arbitrarily large images, removing the need for tiling and its boundary artifacts.
  • The list domain provides a natural interface for object manipulation: editing positions, swapping features, or adding or removing entries before cycling back to an image, which could support data augmentation and interactive editing.
  • The success on low-contrast microscopy cells suggests GAN-based cycle-consistency is a viable alternative to autoencoder slot models for biological imaging, where objects are numerous and visually similar.
  • The method's requirement for a list domain could be extended to other structured outputs (e.g., keypoints, text descriptions) as long as a differentiable cycle can be defined.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The auxiliary scorer target t_p in Eq. (3) is a potential leakage of ground-truth object locations; if it is computed from labels, the method is not fully unsupervised, and if it is recomputed from the model's own predictions, the cycle loss could be trivially satisfied. The paper does not resolve this ambiguity, and it is the main open question for reproducing the unsupervised claim.
  • Because the list cycle loss relies on a discrete assignment, the method may be sensitive to the top-k hyperparameter; inference suggests that a mismatch between k and the true object count could degrade detection, and a learned or adaptive k would be a natural extension.
  • The method's generality beyond homogeneous-background scenes is untested; given the paper's own limitation about complex backgrounds, a testable extension is to combine ORGAN with a background model or to evaluate on datasets with textured backgrounds.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces ORGAN, a CycleGAN-based architecture for object-centric representation learning that maps between images and 'lists' (object locations, presence probabilities, and feature vectors). It reports competitive F1 scores on synthetic datasets (Tetrominoes, Sprites, Multi-MNIST) and claims to be the only method tested that successfully detects objects on a real-world low-contrast Cells dataset, while also scaling to larger images. Comparisons are made against SPACE, SLATE, LSD, and SPOT, with additional experiments on latent space structure.

Significance. If the claims hold, ORGAN would be a novel GAN-based alternative to autoencoder/slot-attention models, particularly valuable for microscopy data with many small objects. The paper includes a reasonable architectural contribution, a broad comparison, and a scaling analysis. However, the central 'unsupervised' claim and the scaling results contain ambiguities and apparent inconsistencies that must be resolved before the contribution can be assessed.

major comments (3)
  1. [Eq. (3), §3.2, §4.2] The list cycle loss in Eq. (3) includes λ_loc ||s_ϕ(p) − t_p||²/|t_p|, where t_p is 'the target for the scorer calculated from the locations of the objects.' For synthetic datasets these locations come from the simulator. For Cells, the paper does not state how t_p or the real list domain is constructed; §4.2 mentions only 50 expert-annotated images for evaluation. If t_p is derived from these annotations during training, ORGAN receives object-location supervision that the baselines do not, invalidating the 'unsupervised' and 'only approach' claims. If t_p is derived from the model's own output, the scorer gradient is circular. The authors must specify the data flow for t_p and the list domain for all datasets.
  2. [§3.3, §4.4] The list generator extracts exactly k patches, with k 'tuned according to the maximum expected number of objects' (§3.1.1). For Sprites, k=10 (§3.3). In the scaling experiment (§4.4), ORGAN is trained on 128×128 Sprites (k=10) and evaluated on 256×256 images with up to 39 objects, yet Fig. 5A reports recall of 0.934 at 39 objects. With k=10, recall cannot exceed 10/39 ≈ 0.256, so either k was increased at inference without disclosure, or the reported recall is inconsistent with the architecture. This also bears on the claim in §1 that 'CycleGANs correctly identify object counts' — the count is upper-bounded by k. Please clarify how k is set in the scaling experiment and discuss the implications.
  3. [§4.2, Table 1] Table 1 reports F1 scores with large standard deviations (e.g., LSD Tetrominoes 20.0±39.8, SPOT 13.8±44.6) but does not state the number of seeds/runs. For Cells, where the main claim rests, ORGAN's F1=77.9±8.6, but the baselines' near-zero scores are plausible; still, the synthetic-dataset 'performs similarly' claim is weakened by variance. Additionally, §4.2 says SPOT's object locations come from its segmentation masks, but Fig. 3 and the caption state SPOT does not allow image reconstruction; explain how masks are obtained.
minor comments (4)
  1. [§4.4] Typo: 'ORGAN outperformed ORGAN' should be 'ORGAN outperformed SPACE'.
  2. [Table 1] Check units/typos: SLATE Tetrominoes training time is 6.75 h while other SLATE entries are 0.75 h; SPACE Sprites training time '4.08' may be '40.8'.
  3. [Availability Statement] The GitHub link lacks a commit hash; the code and dataset pipeline (especially list construction for Cells) are not reproducible from the information given.
  4. [§3.3, §5] The paper's framing as 'unsupervised' should be tempered given that ORGAN explicitly requires a list domain. For real-world datasets, deriving lists typically requires object-location information, which is a form of supervision unless the paper explains otherwise.

Circularity Check

2 steps flagged

ORGAN's detection result is driven by Eq. (3)'s t_p term computed from object locations and by the preset list size k, so the 'unsupervised / only approach' claims reduce to the method's own inputs.

specific steps
  1. fitted input called prediction [Eq. (3), §3.2; §4.2; §5]
    "LCyc L =E L(1/k Σ_l [CObj(l,lCyc) + ||ηl−ηCyc l||^2/|ηl|] + λ_loc ||sϕ(p)−tp||^2/|tp|), with tp being the target for the scorer calculated from the locations of the objects. ... ORGAN requires a dataset that contains lists, where the size is controlled by a hyperparameter."

    The scorer sϕ is trained with a direct regression term λ_loc||sϕ(p)−tp||^2/|tp| whose target is 'calculated from the locations of the objects.' For synthetic data those locations are simulator ground truth; for Cells the only ground-truth locations mentioned are the 50 expert-annotated images in §4.2, and §5 concedes that ORGAN requires a dataset that contains lists. Thus the predicted patch scores and therefore the reported object locations are fitted to the supplied list locations, and the F1/recall numbers measure that fit rather than an unsupervised discovery. The paper's 'unsupervised' and 'only approach tested here capable' claims thereby reduce to the availability of a location-annotated list domain. If instead t_p were derived from the model's own output, the gradient would be circ

  2. self definitional [Intro, §3.1.1, §5]
    "we show that CycleGANs correctly identify object counts even when test-time properties, such as object spacing and object count, differ from the training data. ... The number of extracted patches k is a hyperparameter that can be tuned according to the maximum expected number of objects within an image. ... ORGAN requires a dataset that contains lists, where the size is controlled by a hyperparameter."

    The output list has exactly k entries, and k is chosen as the maximum expected number of objects. The datasets are defined with lists of fixed size (Tetrominoes: three entries; Sprites, Multi-MNIST, Cells: ten entries). When ORGAN is evaluated on 256×256 images 'with up to 39 objects,' k must be set to that maximum, so the object count is an input rather than a predicted quantity. The advertised ability to 'correctly identify object counts' is therefore the prescribed list size by construction, not an emergent count-estimation capability.

full rationale

The core CycleGAN image/list cycle, the GAN losses, the Hungarian assignment, and the feature-space experiments are defined independently and are not circular by themselves. There is no load-bearing self-citation chain: Ihle et al. (2019) is used only as background for CycleGAN segmentation. However, the central detection claim is undermined by Eq. (3), where the scorer is trained against t_p 'calculated from the locations of the objects.' For a real-world dataset, the paper never explains how the list domain is obtained, and §5 admits that ORGAN requires a dataset that contains lists; the only cell ground truth mentioned is expert annotation. If those locations provide t_p, ORGAN's detection is supervised rather than unsupervised; if t_p is model-derived, the cycle-loss gradient is circular. Additionally, the fixed hyperparameter k prescribes the output list size, so the 'object count' generalization claim reduces to the chosen k. Together these two issues make the headline claims—unsupervised, only approach capable on low-contrast real data, and scaling with object count—depend on the method's own inputs. I therefore set the circularity score at 6: partial circularity in the central claim, while other components retain independent content.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The ledger is dominated by hyperparameters (k, patch size, sigma, loss weights) and two assumptions: the list domain supplies object locations, and blob-based reconstruction is expressive enough. No new physical entities are introduced.

free parameters (5)
  • k (number of list entries / top-k) = 3 (Tetrominoes), 10 (Sprites/Multi-MNIST/Cells)
    Chosen to match maximum number of objects per dataset; fixes output list size and therefore bounds the detectable object count.
  • Patch size ph=pw and stride = per-dataset; stride = pw/8
    Adjusted according to dataset; must fit the largest object. Affects detection of objects crossing patch boundaries.
  • Gaussian blob sigma = pw/10
    Fixed to patch size; determines object size in reconstruction and matching cost.
  • Loss weights lambda_DisL, lambda_DisI, lambda_CycI, lambda_CycL, lambda_pres, lambda_loc = not reported
    Weighting of adversarial and cycle losses; no values or tuning procedure given.
  • Feature dimensionality |eta| = dataset-dependent
    Set per dataset (e.g., 3 for Sprites) to represent object properties.
axioms (4)
  • domain assumption Cycle-consistent translation between image and list domains is a well-posed training signal for object discovery.
    The whole method relies on CycleGAN's unpaired translation assumption; if list and image domains are not aligned in distribution, cycle loss can be satisfied by ignoring objects.
  • ad hoc to paper Object locations are available for the scorer target t_p in Eq. (3).
    The loss uses "target for the scorer calculated from the locations of the objects"; for unlabeled Cells images this is not explained and may constitute supervision.
  • standard math Differentiable top-k with non-maximum suppression provides usable gradients for selection.
    Relies on Cordonnier et al. top-k; not proved in the paper, but is an established technique.
  • domain assumption Objects can be represented as fixed isotropic Gaussian blobs plus a U-Net style transfer.
    Borrowed from BlobGAN; fails when objects are larger than patches or too close together.

pith-pipeline@v1.3.0-alltime-deepseek · 12567 in / 12193 out tokens · 111140 ms · 2026-08-02T19:27:00.109640+00:00 · methodology

0 comments
read the original abstract

Although data generation is often straightforward, extracting information from data is more difficult. Object-centric representation learning can extract information from images in an unsupervised manner. It does so by segmenting an image into its subcomponents: the objects. Each object is then represented in a low-dimensional latent space that can be used for downstream processing. Object-centric representation learning is dominated by autoencoder architectures (AEs). Here, we present ORGAN, a novel approach for object-centric representation learning, which is based on cycle-consistent Generative Adversarial Networks instead. We show that it performs similarly to other state-of-the-art approaches on synthetic datasets, while at the same time being the only approach tested here capable of handling more challenging real-world datasets with many objects and low visual contrast. Complementing these results, ORGAN creates expressive latent space representations that allow for object manipulation. Finally, we show that ORGAN scales well both with respect to the number of objects and the size of the images, giving it a unique edge over current state-of-the-art approaches.

Figures

Figures reproduced from arXiv: 2603.02063 by Ellen van Maren, Jo\"el K\"uchler, Katarina Vuli\'c, Reza Abbasi-Asl, Stephan J. Ihle, Vaiva Vasiliauskait\.e.

Figure 1
Figure 1. Figure 1: Model architecture and training losses. ORGAN is based on the Cycle-Consistent Generative Adversarial [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Network architectures used in this work. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of object detection performance for four different datasets. We show performance for ORGAN [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Latent space quality of ORGAN. (A) The latent space as encoded by η is presented for the Sprites dataset. Two of the three dimensions are plotted. The third dimension is kept fixed at its center. The feature space encodes the color, shape, and size of the object. (B) An input image of the Tetrominoes dataset is cycled, while the list elements are modified. Top: All objects were moved closer to the center. … view at source ↗
Figure 5
Figure 5. Figure 5: Analysis of the generalization capabilities of the object detection accuracy. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 17 linked inside Pith

  1. [1]

    Differentiable mathematical programming for object-centric representation learning.arXiv preprint arXiv:2210.02159,

    Adeel Pervez, Phillip Lippe, and Efstratios Gavves. Differentiable mathematical programming for object-centric representation learning.arXiv preprint arXiv:2210.02159,

    Pith/arXiv arXiv

  2. [8]

    Adversarial synthesis learning enables segmentation without target modality ground truth

    Yuankai Huo, Zhoubing Xu, Shunxing Bao, Albert Assad, Richard G Abramson, and Bennett A Landman. Adversarial synthesis learning enables segmentation without target modality ground truth. In2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 1217–1220. IEEE,

    2018

  3. [12]

    Data synthesis based on generative adversarial networks.arXiv preprint arXiv:1806.03384,

    Noseong Park, Mahmoud Mohammadi, Kshitij Gorde, Sushil Jajodia, Hongkyu Park, and Youngmin Kim. Data synthesis based on generative adversarial networks.arXiv preprint arXiv:1806.03384,

    Pith/arXiv arXiv

  4. [13]

    Cyclegan-vc2: Improved cyclegan-based non-parallel voice conversion

    Takuhiro Kaneko, Hirokazu Kameoka, Kou Tanaka, and Nobukatsu Hojo. Cyclegan-vc2: Improved cyclegan-based non-parallel voice conversion. InICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6820–6824. IEEE,

    2019

  5. [14]

    (un) paired signal-to-signal translation with 1d conditional gans.arXiv preprint arXiv:2403.04800,

    Eric Easthope. (un) paired signal-to-signal translation with 1d conditional gans.arXiv preprint arXiv:2403.04800,

    Pith/arXiv arXiv

  6. [15]

    Text-to-image-to-text translation using cycle consistent adversarial networks

    Satya Krishna Gorti and Jeremy Ma. Text-to-image-to-text translation using cycle consistent adversarial networks. arXiv preprint arXiv:1808.04538,

    Pith/arXiv arXiv

  7. [16]

    Object discovery with a copy-pasting gan.arXiv preprint arXiv:1905.11369,

    Relja Arandjelovi´c and Andrew Zisserman. Object discovery with a copy-pasting gan.arXiv preprint arXiv:1905.11369,

    Pith/arXiv arXiv 1905

  8. [18]

    Ssd: Single shot multibox detector

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. InComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14, pages 21–37. Springer,

    2016

  9. [19]

    Yolov3: An incremental improvement.arXiv preprint arXiv:1804.02767,

    Joseph Redmon. Yolov3: An incremental improvement.arXiv preprint arXiv:1804.02767,

    Pith/arXiv arXiv

  10. [20]

    Illiterate dall-e learns to compose.arXiv preprint arXiv:2110.11405,

    Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dall-e learns to compose.arXiv preprint arXiv:2110.11405,

    Pith/arXiv arXiv

  11. [23]

    Sindy Löwe, Phillip Lippe, Maja Rudolph, and Max Welling

    URL https://arxiv.org/abs/2410.00728. Sindy Löwe, Phillip Lippe, Maja Rudolph, and Max Welling. Complex-valued autoencoders for object discovery.arXiv preprint arXiv:2204.02075,

    Pith/arXiv arXiv

  12. [24]

    U-net: Convolutional networks for biomedical image segmenta- tion

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmenta- tion. InMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234–241. Springer,

    2015

  13. [25]

    Spectral normalization for generative adversarial networks.arXiv preprint arXiv:1802.05957,

    Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks.arXiv preprint arXiv:1802.05957,

    Pith/arXiv arXiv

  14. [1995]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    Diederik P Kingma. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    Pith/arXiv arXiv

  15. [2010]

    Movi: A large multipurpose motion and video dataset

    Saeed Ghorbani, K Mahdaviani, Anne Thaler, K Kording, DJ Cook, G Blohm, and NF Troje. Movi: A large multipurpose motion and video dataset. arxiv 2020.arXiv preprint arXiv:2003.01888, 9,

    Pith/arXiv arXiv 2020

  16. [2012]

    Blood dataset

    Irfan Bin Azmi. Blood dataset. 2024-07-08. url:.https://www.kaggle.com/datasets/irfanbinazmi/blood-dataset,

    2024

  17. [2014]

    Contrastive learning of structured world models.arXiv preprint arXiv:1911.12247,

    Thomas Kipf, Elise Van der Pol, and Max Welling. Contrastive learning of structured world models.arXiv preprint arXiv:1911.12247,

    Pith/arXiv arXiv 1911

  18. [2016]

    Gmair: Unsupervised object detection based on spatial attention and gaussian mixture model.Computational Intelligence and Neuroscience, 2022(1): 7254462,

    Weijin Zhu, Yao Shen, Mingqian Liu, and Lizeth Patricia Aguirre Sanchez. Gmair: Unsupervised object detection based on spatial attention and gaussian mixture model.Computational Intelligence and Neuroscience, 2022(1): 7254462,

    2022

  19. [2018]

    Semantic-aware generative adversarial nets for unsupervised domain adaptation in chest x-ray segmentation

    Cheng Chen, Qi Dou, Hao Chen, and Pheng-Ann Heng. Semantic-aware generative adversarial nets for unsupervised domain adaptation in chest x-ray segmentation. InMachine Learning in Medical Imaging: 9th International Workshop, MLMI 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16, 2018, Proceedings 9, pages 143–151. Springer,

    2018

  20. [2019]

    Benchmarking and analysis of unsupervised object segmentation from real-world single images.International Journal of Computer Vision, 132(6):2077–2113,

    Yafei Yang and Bo Yang. Benchmarking and analysis of unsupervised object segmentation from real-world single images.International Journal of Computer Vision, 132(6):2077–2113,

    2077

  21. [2020]

    Synthesizing tabular data using generative adversarial networks.arXiv preprint arXiv:1811.11264,

    Lei Xu and Kalyan Veeramachaneni. Synthesizing tabular data using generative adversarial networks.arXiv preprint arXiv:1811.11264,

    Pith/arXiv arXiv

  22. [2021]

    Bridging the gap to real-world object-centric learning

    Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Schölkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860,

    Pith/arXiv arXiv

  23. [2022]

    On the binding problem in artificial neural networks

    Klaus Greff, Sjoerd Van Steenkiste, and Jürgen Schmidhuber. On the binding problem in artificial neural networks. arXiv preprint arXiv:2012.05208,

    Pith/arXiv arXiv 2012

  24. [2023]

    Tianyang Chen, Yufei Huang, Zhiqiang Shen, Jian Zhang, Yao Li, and Lei Zhang

    URL https://arxiv.org/ abs/2303.10834. Tianyang Chen, Yufei Huang, Zhiqiang Shen, Jian Zhang, Yao Li, and Lei Zhang. Learning global object-centric representations via disentangled slot attention.Machine Learning, 114(40):1–22,

    Pith/arXiv arXiv

  25. [2024]

    Space: Unsupervised object-oriented scene representation via spatial attention and decomposition.arXiv preprint arXiv:2001.02407,

    Zhixuan Lin, Yi-Fu Wu, Skand Vishwanath Peri, Weihao Sun, Gautam Singh, Fei Deng, Jindong Jiang, and Sungjin Ahn. Space: Unsupervised object-oriented scene representation via spatial attention and decomposition.arXiv preprint arXiv:2001.02407,

    Pith/arXiv arXiv 2001

  26. [2025]

    Karl Stelzner, Robert Peharz, and Kristian Kersting

    doi:10.1007/s10994-024- 06687-9. Karl Stelzner, Robert Peharz, and Kristian Kersting. Faster attend-infer-repeat with tractable probabilistic models. In International Conference on Machine Learning, pages 5966–5975. PMLR,

    doi:10.1007/s10994-024-