Source-linked AI summary
Fast Camouflaged Object Detection via Edge-based Reversible Re-calibration Network
Ge-Peng Ji, Lei Zhu, Mingchen Zhuge, Keren Fu
TL;DR
Camouflaged object detection is challenging because foreground objects resemble their surroundings and have ambiguous, weak boundaries. ERRNet addresses this with edge-based re-calibration, achieving top performance across COD and medical segmentation benchmarks while maintaining high inference speed.
Problem
Camouflaged object detection must determine object locations and weak boundaries when foreground and background patterns are highly similar.
Method
ERRNet combines Selective Edge Aggregation and a Reversible Re-calibration Unit to model visual perception, exploit edge priors, and compare potential regions with background.
Results
ERRNet achieves 1st place on three COD datasets, outperforms cutting-edge models on five medical image segmentation datasets, and reaches 79.3 FPS with a 0.867 E-measure.
Takeaways & Limitations
The results support ERRNet as a general and robust solution for camouflaged object detection across COD and medical image segmentation tasks.
Takeaways & Limitations
Performance degrades substantially when foreground and background have strong similarity, and the method relies on low-level patterns such as texture, gradient, and colour information.
Abstract
from arXiv · showhide
Camouflaged Object Detection (COD) aims to detect objects with similar patterns (e.g., texture, intensity, colour, etc) to their surroundings, and recently has attracted growing research interest. As camouflaged objects often present very ambiguous boundaries, how to determine object locations as well as their weak boundaries is challenging and also the key to this task. Inspired by the biological visual perception process when a human observer discovers camouflaged objects, this paper proposes a novel edge-based reversible re-calibration network called ERRNet. Our model is characterized by two innovative designs, namely Selective Edge Aggregation (SEA) and Reversible Re-calibration Unit (RRU), which aim to model the visual perception behaviour and achieve effective edge prior and cross-comparison between potential camouflaged regions and background. More importantly, RRU incorporates diverse priors with more comprehensive information comparing to existing COD models. Experimental results show that ERRNet outperforms existing cutting-edge baselines on three COD datasets and five medical image segmentation datasets. Especially, compared with the existing top-1 model SINet, ERRNet significantly improves the performance by $\sim$6% (mean E-measure) with notably high speed (79.3 FPS), showing that ERRNet could be a general and robust solution for the COD task.
1. Introduction
COD is difficult because camouflaged objects resemble their surroundings and have ambiguous boundaries. ERRNet models edge cues, global guidance, and cross-comparison through SEA and RRU, achieving strong accuracy and fast inference.
- Motivation: COD is challenging because objects vary in size, have ambiguous boundaries, and resemble their environments in texture.
- Results: ERRNet achieves competitive performance and faster inference than state-of-the-art COD methods, with real-time inference across eight challenging datasets.
- Motivation: Human-inspired COD should combine weak boundary cues, global guidance for proposing regions, and cross-comparison with surrounding areas.
- Motivation: Existing COD models seldom completely model this biologically inspired process, leading to less satisfactory results.
- Method: ERRNet uses Selective Edge Aggregation (SEA) and a Reversible Re-calibration Unit (RRU) to model edge priors and cross-comparison between objects and backgrounds.
- Results: 0.867 E-measure and 79.3 FPS demonstrate ERRNet’s reported accuracy–speed combination, while SINet achieves 5 FPS with 6% lower accuracy.
- Method: RRU re-calibrates coarse inference maps using Neighbour, Global, Edge, and Semantic priors to contrast potential objects with complementary areas.
2. Related Work
Related work spans hand-crafted camouflage detection, CNN-based COD, and edge-aware binary segmentation. The paper positions ERRNet as a scheme for more accurate localization and boundary delineation using diversified priors.
- Camouflaged Object Detectors: Early camouflage detectors used hand-engineered features to compare foreground regions with their environments.These approaches included co-occurrence matrices, optical flow, spatial-colour GMMs, and 3D convexity-based methods.
- Camouflaged Object Detectors: Hand-crafted methods work well only in limited cases and degrade when foreground and background are strongly similar.The cited limitation is especially relevant to camouflage because object and environment patterns can closely resemble each other.
- Camouflaged Object Detectors: CNN-based COD methods introduced annotated datasets and architectures that combine classification or search modules with segmentation or identification streams.Examples include CAMO with classification and segmentation streams, and SINet with search and identification modules.
- Edge-aware Binary Image Segmentation: Edge-aware segmentation research learns edge cues, edge losses, or edge branches to improve salient-object boundaries and details.Reported designs include IoU edge loss, boundary-aware loss, hybrid losses, edge detection branches, and cross-refinement units.
- Proposed distinction: ERRNet differs from prior work by incorporating diversified priors in the decoder to improve camouflaged-object localization and boundary delineation.The paper states that these priors promote COD performance.
3. Proposed Framework
ERRNet uses a ResNet-50 encoder and cascaded decoder components to combine global, edge, neighbour, and semantic priors. SEA supplies selectively aggregated edge cues, while RRU reversibly recalibrates coarse predictions by comparing foreground and complementary regions.
- Overall framework: ERRNet uses ASPP on the encoder’s highest feature hierarchy to enlarge the receptive field and fuse multi-context information.The ASPP adopts dilation rates d ∈ {1, 6, 12, 18}.
- Reversible Re-calibration Unit: RRU combines NGES priors to recalibrate coarse predictions by contrasting potential camouflaged regions with their complement areas.The module performs progressive refinement with diversified priors in a cascaded manner.
- Prior construction: NGES priors comprise neighbour prior maps, edge prior features, global prior features from ASPP, and semantic prior features from the encoder.These priors are fed to each RRU and provide complementary cues for re-calibration.
- Selective Edge Aggregation: SEA adaptively selects mutual representations from low-level edge-aware features before aggregation and applies explicit edge supervision.The resulting edge prior is passed to the re-calibration stage.
- Reversible Re-calibration Unit: RRU reverses estimated camouflaged regions to mine complementary details while edge priors help locate camouflaged boundaries.The mechanism refines coarse maps in a cascaded fashion using bottom-up edge guidance and top-down global guidance.
- Prior effects: The prior visualizations associate edge priors with fine-grained weak boundaries, global priors with region localization, and neighbour priors with prediction stability.The paper states that none of the NGES priors is dispensable.
4. Experiments
The experiments introduce benchmark datasets and evaluation metrics, then assess ERRNet’s effectiveness. The paper also states that trained models, benchmark results, code, and application results will be released.
- Experimental setup: The experiments use benchmark datasets and evaluation metrics to verify the effectiveness of ERRNet.The supplied passage introduces the experimental evaluation without reporting dataset names or metric values.
- Reproducibility: The authors plan to release trained models, COD benchmark results, code, and results from two applications upon publication.This is presented as a release plan rather than an experimental finding.
this work. 2
ERRNet is evaluated on three COD benchmarks and through ablations of its priors, reversible operation, edge aggregation, and multi-context aggregation. It achieves strong detection performance alongside 79.3 FPS inference speed.
- Datasets and setup: ERRNet is evaluated on COD10K, CAMO, and CHAM, using combined COD10K and CAMO training data and separate benchmark test sets.COD10K contains 3,040 training and 2,026 testing images; CAMO contains 1,000 training and 250 testing images; CHAM contains 76 testing images.
- Ablation study: ERRNet outperforms EAA and EAM across all four metrics, indicating that SEA learns edge information better than addition or multiplication fusion.The ablation compares the proposed Selective Edge Aggregation against Edge Aggregation via Addition and Edge Aggregation via Multiplication.
- Datasets and setup: 79.3 FPS is achieved at 352 × 352 input resolution without post-processing such as CRF.The implementation uses an RTX TITAN GPU for accelerated experiments.
- Ablation study: Reverse operation improves scores over the w/o-reverse variant, supporting reversible attention for identifying camouflaged objects.The RRU first predicts a camouflaged detection map and reverses it to obtain a reverse attention map.
- Ablation study: Combining global, neighbour, and edge priors in RRU produces better COD performance than the first three ablated variants.The full method achieves superior best metric results over variants #1, #2, and #3.
- Quantitative comparisons: ERRNet outperforms 13 compared SOTA methods across all three COD datasets and more accurately detects camouflaged objects.The paper reports superior best metric results over the compared methods.
- Efficiency comparison: 79.3 FPS and Eφ = 0.867 exceed CPD’s 62 FPS and Eφ = 0.770, with a reported 9.7% performance improvement.The reported speed gap is 27.9%.
5. Conclusion
The paper presents ERRNet, combining SEA and RRUs with multiple priors for camouflaged object detection. It reports first-place COD results and strong transfer to medical image segmentation, while identifying additional cues and modalities for future work.
- Proposed network: ERRNet combines Selective Edge Aggregation modules and Reversible Re-calibration Units with neighbour, global, edge, and semantic priors.SEA mines edge information and RRU re-calibrates coarse predictions using complementary priors.
- Results: ERRNet achieves first place on three COD datasets and outperforms cutting-edge models on five medical image segmentation datasets.The conclusion presents ERRNet as a general and robust solution for COD.
- Future work: The authors plan to explore texture, gradient, and colour cues as auxiliary information for deep COD models.These cues are proposed as additional low-level patterns alongside object edges.
- Future work: Future work also considers depth or thermal infrared modalities to improve detection accuracy in related RGB-D and RGB-T tasks.The paper explicitly identifies these modalities as a planned extension.