acceptodds
Withdrawn from ICLR 2027

Learning Video Camouflaged Object Detection from Unlabeled Static Images

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Video Camouflaged Object Detection (VCOD) is a criuital security technology for identifying camouflaged objects in videos. Existing VCOD methods often rely on motion signals across consecutive frames to capture dynamic cues. However, these motion signals are inherently noisy and inaccurate, and excessive reliance on them can lead to error accumulation. Moreover, the lack of camouflaged video datasets and the costly annotation process hinder progress in this field. To address these challenges, we propose SDC, a static-dynamic correspondence framework that extracts pseudo-dynamic signals from unlabeled static images, effectively capturing static-dynamic correspondences. Initially, we leverage coordinate data from the contrastive crop views to establish static correspondences, ensuring the consistency of static feature representations. Following that, we design a dynamic capture layer that generates pseudo-dynamic signals in both forward and backward directions. These pseudo-dynamic signals serve as cues for learning dynamic representations. Finally, we introduce a static-dynamic consistency loss function to enforce consistency between static and dynamic correspondences. Notably, our method requires only unlabeled static image data for training, freeing VCOD from its dependency on video data. By using only static images for training, SDC outperforms existing video-dependent strategies, surpassing state-of-the-art weakly supervised, unsupervised, and even surpasses several fully supervised video methods on public benchmarks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Video, Camouflaged object detection, Static images, Pseudo-dynamic signals, Unsupervised Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sdJoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SDC, an unsupervised framework for VCOD trained exclusively on unlabeled static images. The method constructs two cropped views of each image and employs online and momentum-updated target encoders to extract visual representations. A pseudo-dynamic module generates bidirectional displacement-like features, while coordinate-based positive sample selection supports joint static and dynamic correspondence learning.The method is trained on static images from COD10K and CAMO and evaluated on MoCA-Mask and CAD2016. The authors report competitive segmentation performance, fast training, and high inference efficiency without requiring video training data.

Critical strengths
  • Practical relevance of annotation-free training.

Reducing dependence on annotated video data is meaningful for VCOD, where collecting and annotating videos can be expensive. The reported results suggest that static-image-based training may offer a computationally efficient alternative.

Critical weaknesses
  • Severe lack of methodological novelty and missing attribution to highly relevant prior work.

My primary concern is the substantial overlap between the proposed SDC framework and the earlier work:

Pei et al., Dynamic in Static: Hybrid Visual Correspondence for Self-Supervised Video Object Segmentation (HVC), arXiv:2404.13505, 2024.

HVC had already introduced the central idea of learning static and pseudo-dynamic correspondences exclusively from unlabeled static images. More importantly, the resemblance extends well beyond the high-level motivation.

SDC's Figure 3 closely resembles HVC's Figure 3 in its overall architecture, including the online/target encoders, EMA updates, projector/predictor structure, crop-coordinate warping, bidirectional pseudo-dynamic generation, positive correspondence masks, static/dynamic similarity matrices, and their combined optimization.

The mathematical formulations are also substantially overlapping. In particular, SDC's bidirectional pseudo-dynamic generation in Eq. (6), coordinate correspondence construction in Eqs. (8)–(10), and final consistency objective in Eq. (11) closely follow the corresponding formulations in HVC.

The main identifiable differences appear to be the application to VCOD, the replacement of the cropping strategy with ContrastiveCrop, distance-weighted positive masks, and backbone or implementation changes. These modifications do not adequately establish a fundamentally new learning framework.

Most concerningly, I could not identify a citation to HVC in the submitted manuscript, despite the substantial methodological and architectural overlap. The lack of attribution and an explicit comparison makes it difficult to assess the actual contribution of SDC. This issue is fundamental to my recommendation.

  • Important methodological ambiguities and inconsistent experimental reporting.

The paper does not sufficiently explain how the learned correspondence representations are converted into final camouflage segmentation masks during inference. The architecture primarily describes self-supervised representation learning, but the complete inference procedure for VCOD remains unclear.

There are also inconsistencies in the reported experiments and implementation. For example, Table 1 reports an S-measure of 0.671 on MoCA-Mask, whereas Table 2 reports 0.724 for the full configuration. The manuscript does not adequately clarify the different evaluation conditions, if any.

Furthermore, Eq. (11) normalizes the correspondence similarities by the sum of positive-mask weights, whereas the Appendix pseudocode uses a mean over the masked similarity matrices. These computations are not generally equivalent.

Questions for authors

Q1. Relationship to HVC. What are the precise methodological differences between SDC and HVC (Pei et al., 2024)? Given the extensive architectural and mathematical similarities, why is HVC not cited or discussed? Which components constitute genuinely new contributions beyond HVC?

Q2. Inference and reproducibility. How are final segmentation masks generated from the learned representations without additional supervision or reference masks? Can the authors clarify the discrepancy between Tables 1 and 2 and identify the exact loss implementation used to produce the reported results?

Reviewer_ifd9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SDC, an unsupervised framework for VCOD that is trained only on unlabeled static images (COD10K + CAMO). The VCOD datasets (MoCA-Mask and CAD2016) are only used for testing. Two overlapping crops are taken from each image. A static consistency loss aligns the features of the two crops, and a small pseudo-dynamic module predicts offset-like signals between them, trained with consistency losses. The authors report results that are better than other unsupervised and weakly supervised methods.

Critical strengths
  • The paper is well written and well organized.
  • The idea of learning VCOD without video data and labels is interesting, and making motion modelling self-supervised through consistency losses is the main contribution in my view.
Critical weaknesses
  • The training is done on static COD images, not on VCOD frames. Then the “pseudo-dynamic” signal is only the crop-and-resize transform between two views of the same image, so the object and the background move together. This is not the motion that helps to find camouflaged objects in videos. Also, at test time the input is full video frames, not two crops of one image. It is not clear if the pseudo-dynamic module is used at test time at all. If each frame is processed independently, then the method is basically unsupervised image COD tested on video frames, and the “video” and “dynamic” claims are not supported.

  • When the authors train on MoCA-Mask frames (Table 2(e)), the result is not better. That is strange: in-domain video data should help, and real adjacent frames are available but not used.

  • The meaning of the offset QQ is not clear. The two crops have different sizes and positions in the original image and are resized to the same size, so the relation between them is a scale and translation change, not optical flow. The paper also says QQ is only “displacement-like” and not physical optical flow. But Table 3 evaluates QQ by EPE against RAFT/PWC-Net flow, as if it were optical flow. If QQ is not flow and has no pixel scale (it is learned with a cosine loss), I am not sure what this EPE means.

  • It is not clear how the final per-pixel mask is produced. The network is trained with a self-supervised feature loss and there is no segmentation head in the paper. But Fig. 4 shows results that are almost the same as GT, even for very thin structures, while FβwF^w_\beta in Table 1 is only 0.444. I am not sure the results are really that good. Also, the paper claims the training is unsupervised on COD10K + CAMO, but Fig. 3 uses a frame from MoCA-Mask as the “static image” example. This is confusing. It is not clear whether any video frames or labels are used.

  • The method feels quite engineering-heavy. Most parts are existing techniques (pixel-level BYOL/PixPro-style loss, ContrastiveCrop, EMA), and they are packaged as a “static-dynamic correspondence” theory. Also, Table 1 and Table 2 give different numbers on MoCA-Mask (0.671 vs 0.724).

Questions for authors
  1. Is the training on COD10K + CAMO fully unsupervised, including any decoder or segmentation head? Fig. 3 shows a MoCA-Mask frame as the input example, which makes it confusing which data is actually used for training.

  2. The training is on static images, not VCOD frames. Then for “This network is responsible for calculating pixel offset vectors between overlapping regions of cropped image views”, you need to assume the overlapping regions are the same object. In a static image this is always true, because object and background have the same transform. How does this help to learn real object motion in videos? At test time, does SDC take multiple frames as input, or each frame independently? Is the pseudo-dynamic module used at inference? If not, how does the learned “dynamic” information help VCOD?

  3. In Table 2(e), when training on MoCA-Mask frames, did you use each frame as a single static image, or did you use real adjacent frames? Why not use the real frame pairs? I think this is the most direct way to show that pseudo-dynamic signals can replace real motion.

  4. What exactly does QQ represent? The two crops have different scales, and QQ is not supervised to match the real displacement between them. Why evaluate it with EPE against RAFT/PWC-Net in Table 3, and how is a pixel-level EPE computed from it? After cropping and resizing, each view loses the global context. How should we understand the offset between two crops? Which coordinate frame is it in, crop or original image?

  5. How do you get the per-pixel segmentation map at test time: a decoder, clustering, or thresholding? Are the examples in Fig. 4 representative? Please show random samples and failure cases.

  6. Why are the MoCA-Mask results different in Table 1 and Table 2? What does T=5T=5 mean?

Reviewer_1Pv1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SDC, a framework for video camouflaged object detection trained only on unlabeled static images. It learns static correspondences from overlapping crops and adds a pseudo-dynamic branch with a consistency loss. Experiments on MoCA-Mask and CAD2016 report competitive accuracy and fast inference without video training data.

Critical strengths
  1. Learning from unlabeled static images is a useful direction that reduces dependence on expensive video annotations.
  2. The reported results are promising under this training setting. The ablations also show gains from adding static and dynamic correspondence learning.
Critical weaknesses
  1. The manuscript does not clearly explain how the learned correspondence features produce the final segmentation masks at inference. This makes the complete annotation-free pipeline difficult to assess and reproduce.
  2. The main and ablation results are inconsistent without a clear explanation. On MoCA-Mask, Table 1 reports S-measure 0.671 and weighted F-measure 0.444, while the full model in Table 2 reports 0.724 and 0.495. The meaning of T = 5 and the evaluation differences need clarification.
Questions for authors
  1. How are the final masks generated from the learned features, and does inference require any initial mask or other supervision?
  2. What explains the difference between Tables 1 and 2, and which configuration supports the main comparison?

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

Related papers

Discussion (0)

Sign in to comment.