acceptodds
Under review as a conference paper at ICLR 2027

CamoWorld: Towards Image World Modeling for Camouflaged Object Understanding

Abstract

Camouflage is a natural concealment mechanism where objects blend into their surroundings and become difficult to distinguish from complex backgrounds. Humans rely on internal world models to infer unobserved states, understand scene regularities, and form plausible expectations from limited observations. In this paper, we present CamoWorld, the first self-supervised latent image world model for camouflaged object understanding. Rather than fitting task-specific outputs or reconstructing pixels, CamoWorld formulates camouflage understanding as latent predictive world modeling. Specifically, it introduces three latent world variables: 1) a mask variable for local structure completion, capturing fine-grained textures, boundaries, and subtle structural discrepancies; 2) a spatial relation variable for global scene continuation, modeling spatial organization and continuity between hidden objects and surrounding backgrounds; and 3) a domain variable for appearance prediction, capturing predictable changes across imaging conditions. CamoWorld learns camouflaged world knowledge from raw images within a unified latent prediction framework. Qualitative and quantitative analyses show its ability to capture such knowledge, while extensive downstream experiments demonstrate that latent world prediction learns more effective representations than existing visual pre-training paradigms.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
World Models, Self-Supervised Learning, Latent Predictive Learning, Visual Representation Learning, Camouflaged Object Understanding, Camouflaged Object Segmentation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_7eikOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper trains a ViT-B on the unlabeled CAMO/COD10K training images with JEPA-style latent prediction, which is a world model for camouflaged objects. Three losses are designed: a) predict masked patches from the visible ones (masks split evenly between textured and flat patches); b) predict a larger crop from a smaller one, given where the small one sits; c) predict the original image from a blurred or brightness/contrast-shifted copy, given the transform parameters. The encoder is then fine-tuned for segmentation, box detection, few-shot and low-label segmentation, and is reported to beat public SSL checkpoints (e.g., I-JEPA, DINOv3) on all of them.

Critical strengths
  1. The paper argues that hidden regions in camouflaged images admit many plausible low-level completions, so predicting latents is a better target than reconstructing pixels. It also address the dominance of flat background patches with a label-free, structure-balanced masking rule. The method is simple and easy to follow, where the three objectives share one encoder and predictor under a single conditioned formulation

  2. The main tables cover segmentation on four benchmarks, box detection on three, few-shot and label-efficient settings, and the method is best on every reported benchmark, with large margins on detection (COD10K-D mAP 36.7 vs 27.9 for DINOv3) and 1-shot segmentation (mIoU 62.8 vs 54.6). Removing any of the three objectives lowers all four metrics (Table 5).

Critical weaknesses
  1. The "pre-training" is small-scale adaptation on the downstream training images. (a) These images are the training images of the downstream benchmarks (labels dropped), so the setting is closer to unsupervised in-domain adaptation than to transferable pre-training. Transfer beyond them is tested only on VOC (+1.2 mIoU over I-JEPA, Table 14b). (b) Pre-training uses 4,040 images, orders of magnitude fewer than the data the compared methods were designed for. Baselines retrained on it (Table 8) run in a regime they were not built for, so Table 8 shows which objective suits 4k images, not which is better in general.

  2. The comparisons seems not support that outperforms existing visual pre-training paradigms or competitiveness with camouflage-specific methods. Table 16 consists mostly of 2018-2022 methods and omits later peer-reviewed ones (e.g., ZoomNet, HitNet, FEDER, FSPNet, CamoFormer, ZoomNeXt) and the NC4K column. C3Net, cited as "(?)", beats CamoWorld by 0.07-0.10 S_α on CAMO and COD10K.

Questions for authors
  1. The Domain objective looks like that of Image World Models (Garrido et al., 2024). How does it differ, and why is IWM not a baseline?

  2. What do the same objectives give when pre-trained on (i) 4,040 non-camouflage images and (ii) a larger unlabeled set disjoint from the downstream training images? This tests the "camouflage" and "pre-training" claims and could raise my Rating.

Reviewer_A1F6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CamoWorld proposes a self-supervised latent image world model for camouflaged object understanding based on a joint-embedding predictive architecture (JEPA). It introduces three complementary objectives: local structural completion through masked latent prediction, global scene continuation through relation-conditioned nested crops, and domain-conditioned latent prediction under appearance transformations. The learned representations are evaluated on camouflaged object segmentation, box-level detection, few-shot transfer, and sample-efficient learning. Extensive experiments, including controlled comparisons under matched pre-training data and budgets, show consistent improvements over representative self-supervised learning baselines.

Critical strengths
  1. A promising modeling perspective. The paper formulates camouflage understanding as conditional latent prediction rather than relying solely on pixel reconstruction or task-specific supervision. The three complementary objectives are designed to capture local structures, global spatial relationships, and appearance variations.

  2. Comprehensive empirical validation. The method is evaluated across segmentation, detection, few-shot transfer, and sample-efficient learning. Importantly, the appendix includes controlled comparisons using matched pre-training data, backbones, and training budgets, strengthening the evidence that the proposed objectives improve downstream representations.

Critical weaknesses
  1. The novelty and scope of the world-model formulation could be clarified. CamoWorld is an image-level latent prediction framework without explicit temporal dynamics. Although its three objectives are reasonable and empirically effective, the distinction between the proposed world-model formulation and existing JEPA-style or conditional latent prediction methods remains insufficiently clear. The paper would benefit from a more precise discussion of its conceptual and methodological contributions.

  2. Some aspects of the objective design and generalization remain underexplored. Global modeling predicts a larger crop that contains the observed context, but it is unclear whether the loss is restricted to the unobserved region or also covers overlapping visible regions. In addition, domain modeling relies on a predefined set of appearance transformations, while evaluation beyond camouflage understanding is limited. Further clarification and analysis would strengthen the claims regarding global structural learning and generalization.

Questions for authors
  1. How sensitive is the local structure scoring function to the relative contributions of gradient magnitude and intensity variation? Could the authors clarify the robustness of this scoring strategy beyond the reported sampling and masking-ratio ablations?

  2. In Global Modeling, is the latent prediction loss computed over the entire target crop, including regions already visible in the context, or only over unobserved regions? How much does the overlap contribute to the observed downstream gains?

  3. Given that CamoWorld does not explicitly model temporal dynamics, how should its world-model formulation be distinguished from general conditional latent prediction methods? What additional challenges would need to be addressed to extend it to dynamic camouflaged video?

Reviewer_9fSzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CamoWorld, a self-supervised latent image world model for camouflaged object understanding. The model follows a joint-embedding predictive architecture with a context encoder, an EMA-updated target encoder, and a predictor, and it defines three latent world variables: a mask variable for local structural completion, a spatial relation variable for local-to-global scene continuation, and a domain variable for appearance prediction under controlled transformations. Pre-training uses only the raw training images of CAMO and COD10K, 4,040 images in total. The paper evaluates transfer on camouflaged object segmentation on CHAMELEON, CAMO, COD10K, and NC4K, box-level detection on COD10K-D, NC4K-D, and CAMO-D, few-shot segmentation on COS-FSS, and sample-efficient learning on COD10K, and reports controlled same-data comparisons, ablations across all four segmentation benchmarks, a three-seed analysis of the domain objective, and PASCAL VOC transfer.

Critical strengths
  1. The controlled same-data and same-budget comparison in Table 8, together with the continued pre-training experiment in Table 9, separates the effect of the proposed objectives from camouflage-data exposure. On COD10K segmentation, CamoWorld improves Fβw from 0.711 to 0.744 over I-JEPA with matched transformations. This is the evidence that most strongly supports the central claim, because the margins over public checkpoints would otherwise be uninterpretable.

  2. The formulation of camouflaged object understanding as latent world modeling is new, and the three-variable design is coherent. The downstream evaluation covers segmentation, detection, few-shot transfer, and sample efficiency, and ablations in Table 15 span all four segmentation benchmarks and show that each objective and each latent variable contributes.

Critical weaknesses
  1. Controlled evidence for the central claim is limited to segmentation. Detection (Table 2), few-shot segmentation (Table 3a), and sample efficiency (Table 3b) compare CamoWorld with public checkpoints of general self-supervised methods that have not seen camouflage images, while CamoWorld is pre-trained on the CAMO and COD10K training images. The margins in these settings, for example 1-shot FB-IoU 80.7 versus 72.5 for DINOv3 and 36.7 versus 27.9 mAP on COD10K-D, are therefore confounded with in-domain exposure. No same-data or continued-pre-training baseline is reported for detection or few-shot, so three of the four downstream claims rest on comparisons where this confound is not removed.

  2. The gains over the closest in-domain baseline are small, and the robustness evidence is thin. Against SSPNet, the only camouflage-domain self-supervised baseline in Table 1, the improvements are 0.005 to 0.008 in Sα and Fβw on CHAMELEON, CAMO, and COD10K, and SSPNet is absent from NC4K, detection, and few-shot. Tables 1 to 3, 5, 8, 9, and 15 report no seed variance; only the Domain analysis in Table 14a uses three seeds, while the Table 5 ablation deltas range from 0.005 to 0.016 in Fβw. The local, global, and domain objectives are also close to existing recipes in self-supervised learning, and although the paper compares Global modeling with multi-crop consistency and Domain modeling with augmentation-based invariance, it does not discuss prior work on guided masking and transformation-conditioned prediction.

Questions for authors
  1. Can you provide same-data or continued-pre-training baselines for detection, few-shot, and sample efficiency, as you did for segmentation in Tables 8 and 9? If the margins in Tables 2 and 3 shrink substantially when I-JEPA or DINOv3 is retrained on the same 4,040 images, I would lower my assessment of the contribution.
  2. Can you report mean and standard deviation over at least three seeds for Tables 1, 2, and 3 and the key ablations in Table 5? Given deltas of 0.005 to 0.016, I need to know whether these differences are stable.
  3. Please clarify the model selection procedure for downstream fine-tuning. Was a validation split used, or were checkpoints and hyperparameters selected on the test sets? If the latter, the reported numbers may be optimistic and I would revise my score downward.
  4. Why is SSPNet evaluated only on segmentation and not on NC4K, detection, or few-shot? Was it run under the same downstream protocol as the other baselines in Table 1?
  5. The structure-guided masking and the transformation-conditioned objective are close in spirit to prior work on guided masking and transformation-aware self-supervision. Can you clarify what is specific to camouflage beyond applying these recipes, and how the formulation differs from that line of work?
  6. Can you add more comparisons, including comparisons among the ablated variants, and more visualization or failure case analysis to explain why the method works? Does the method address problems that are specific to this domain, such as weak boundaries and low contrast, or would the same three objectives work equally well on generic natural images?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.