Source-linked AI summary
Uncertainty-aware Joint Salient Object and Camouflaged Object Detection
Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, Yuchao Dai
TL;DR
SOD and COD address contrasting visual targets, yet their datasets contain partially related examples and uncertain annotations. The paper jointly learns both tasks using cross-task data interaction, contradiction-aware similarity modeling, and adversarial confidence estimation. Experiments on benchmark datasets report state-of-the-art performance for both tasks.
Problem
SOD and COD target contrasting objects, but their datasets are partially related and contain labeling uncertainty that joint modeling can address.
Method
The framework uses easy COD samples as hard SOD samples, a similarity measure for contradictory task attributes, and adversarial learning for prediction confidence.
Results
Experiments on six benchmark SOD datasets and three benchmark COD datasets demonstrate the effectiveness of the joint learning solution.
Takeaways & Limitations
Jointly leveraging contradictory SOD and COD information supports robust saliency detection, camouflage detection, and interpretable confidence maps.
Takeaways & Limitations
The adversarial-learning discriminator and loss weight are selected empirically, leaving more effective alternatives as future work.
Abstract
from arXiv · showhide
Visual salient object detection (SOD) aims at finding the salient object(s) that attract human attention, while camouflaged object detection (COD) on the contrary intends to discover the camouflaged object(s) that hidden in the surrounding. In this paper, we propose a paradigm of leveraging the contradictory information to enhance the detection ability of both salient object detection and camouflaged object detection. We start by exploiting the easy positive samples in the COD dataset to serve as hard positive samples in the SOD task to improve the robustness of the SOD model. Then, we introduce a similarity measure module to explicitly model the contradicting attributes of these two tasks. Furthermore, considering the uncertainty of labeling in both tasks' datasets, we propose an adversarial learning network to achieve both higher order similarity measure and network confidence estimation. Experimental results on benchmark datasets demonstrate that our solution leads to state-of-the-art (SOTA) performance for both tasks.
1. Introduction
SOD and COD target visually opposing but partially overlapping objects, motivating a joint framework that uses their complementary and contradictory information. The proposed method combines cross-task data interaction, similarity modeling, and uncertainty estimation.
- SOD seeks high-contrast objects, whereas COD targets objects sharing structure or texture with their surroundings.
- Objects can be both salient and camouflaged, indicating a partial positive relationship between the two tasks at the dataset level.
- Easy COD samples, such as polar bears, are used as hard positive samples for SOD to provide context-aware contrast-level data augmentation.
- A similarity measure module uses connection-modeling images to force SOD and COD to focus on different image regions despite their contradictory attributes.
- Joint training can transfer precise salient-object local-context information to COD, where object boundaries are difficult to model.
- Adversarial training models confidence because SOD labels can be subjective and COD annotations can be incomplete around visually similar surroundings.
- The framework’s contributions are joint uncertainty-aware SOD/COD learning, explicit contradiction modeling, and COD-to-SOD data interaction for robust saliency detection.
2. Related Work
Prior work largely develops task-specific saliency or camouflage models and often exploits positively related tasks. This paper instead connects contradictory SOD and COD tasks through explicit similarity modeling in a joint framework.
- Salient Object Detection Models: Existing saliency models emphasize structure-preserving predictions through auxiliary edge branches, structure-aware losses, cascade decoding, or boundary-aware mechanisms.
- Salient Object Detection Models: Some saliency methods use auxiliary sources such as fixation prediction or weakly supervised semantic segmentation to support fully supervised SOD.
- Camouflaged Object Detection Models: COD models target objects that reduce conspicuousness through context-dependent mechanisms including background-pattern matching and disruptive coloration.
- The proposed framework contains task-specific feature encoders, a shared prediction decoder, and a shared fully convolutional discriminator for pixel-wise confidence estimation.
- Existing multi-task frameworks mainly exploit positively related tasks, whereas this work explicitly models the contradictory attributes of SOD and COD.
3. Our Method
The method jointly learns salient and camouflaged object detection by interacting their data, modeling their contradictory features, and estimating prediction confidence under labeling uncertainty.
- Framework overview: The framework jointly learns SOD and COD with data interaction, a similarity measure module, and uncertainty-aware adversarial learning.It uses COD samples, a connection-modeling dataset, task-specific encoders and decoders, and a discriminator-based confidence estimation module.
- Data interaction as data augmentation: Easy COD samples with the smallest MAE under a trained SOD model are selected as hard positive samples for SOD augmentation.The method selects M = 400 COD samples and replaces M randomly selected SOD training samples.
- Contradicting modeling: The similarity measure module uses PASCAL VOC 2007 connection-modeling images to separate saliency and camouflage feature representations.The two encoders extract task-specific features, which are concatenated and mapped to latent saliency and camouflage features.
- Contradicting modeling: Figure 4 indicates that the saliency encoder emphasizes contextually conspicuous regions, whereas the camouflage encoder focuses on hidden regions resembling their backgrounds.The two branches process the same image and highlight different foreground regions.
- Uncertainty-aware adversarial learning: The uncertainty-aware adversarial network models task-specific prediction confidence because SOD labels are ambiguous and COD annotations are difficult to complete accurately.A fully convolutional discriminator evaluates decoder predictions and supports adversarial learning for confidence estimation.
- Prediction decoder: The prediction decoder combines task-specific features with lower-level features using residual channel attention, dual attention, and holistic attention modules.The decoder produces detailed saliency and camouflage maps from the respective encoder features.
4. Experimental Results
Experiments evaluate the joint model on six SOD and three COD benchmark datasets, comparing against established models and staged ablations. The model achieves strong performance across both tasks while producing uncertainty estimates, though adversarial training remains difficult to optimize.
- Experimental setup: The experiments train on augmented DUTS for SOD and COD10K for COD, then evaluate across six SOD and three COD testing datasets.The SOD datasets include DUTS, ECSSD, DUT, HKU-IS, THUR, and SOC; COD datasets include CAMO, CHAMELEON, and COD10K.
- SOD performance comparison: The SOD branch achieves 5/6 best performance against eleven state-of-the-art SOD models, except on the SOC testing dataset.Texture images in SOC may be treated as camouflaged objects and influence performance.
- COD performance comparison: The camouflage model consistently performs best among re-trained saliency models evaluated on the COD testing set, supporting the joint learning framework.The uncertainty map represents model confidence toward the current prediction and supports interpretable downstream use.
- Ablation study: Joint training improves both tasks over separately trained ASOD and SCOD, while similarity measurement especially improves COD10K performance.The uncertainty-aware variants JSOD3 and JCOD3 show relatively comparable performance, attributed to the difficulty of training the adversarial branch.
- Ablation study: Adversarial learning uses a small adversarial-loss weight of 0.01 to balance model performance against effective uncertainty estimation.The authors identify searching for a more effective discriminator and loss weight as future work.
5. Conclusion
The paper proposes an uncertainty-aware joint network for salient and camouflaged object detection. It combines cross-task data interaction, similarity modeling, and adversarial confidence estimation, and reports effectiveness across nine benchmark datasets.
- Conclusion: The proposed framework uses easy COD samples as hard SOD samples to learn a more robust saliency model.This data interaction strategy is presented as the first component of the joint solution.
- Conclusion: A similarity measure module explicitly models the contradicting attributes of SOD and COD using an extra connection modeling dataset.The module builds the task connection within the joint framework.
- Conclusion: An adversarial learning network models prediction confidence to address uncertainty in SOD and COD annotations.The framework targets uncertainty arising in both task datasets.
- Conclusion: Experiments on six benchmark SOD datasets and three benchmark COD datasets demonstrate the effectiveness of the joint learning solution.The conclusion reports effectiveness across both detection tasks.