PaperlayerView exact revision on arXiv

Source-linked AI summary

Zoom In and Out: A Mixed-scale Triplet Network for Camouflaged Object Detection

Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, Huchuan Lu

arXiv:2203.02688v1cs.CV

TL;DR

Camouflaged object detection must segment targets that resemble their surroundings and vary in scale, appearance, and occlusion. ZoomNet uses a mixed-scale triplet architecture with SIUs, HMUs, and uncertainty-aware loss to address localization and prediction ambiguity. The paper reports superior performance across COD and SOD benchmarks, while noting slower inference than the fastest comparison method.

  • Problem

    Camouflaged object detection remains difficult because targets are visually insignificant, diverse in scale and appearance, occluded, and easily confused with background surroundings.

  • Method

    ZoomNet uses shared triplet feature extraction across 1.0×, 1.5×, and 0.5× inputs, SIUs and HMUs for mixed-scale representation, and UAL alongside BCE for reliable prediction.

  • Results

    ZoomNet surpasses 23 state-of-the-art methods on four COD datasets under seven metrics and shows superior performance and generalization on SOD.

  • Takeaways & Limitations

    Mixed-scale feature integration and uncertainty-aware supervision support accurate and reliable segmentation across both COD and SOD tasks.

  • Takeaways & Limitations

    The shared, explicit multi-scale extraction design incurs additional inference cost, leaving ZoomNet slightly slower than C2FNet, the fastest method.

Abstract

from arXiv · show

The recently proposed camouflaged object detection (COD) attempts to segment objects that are visually blended into their surroundings, which is extremely complex and difficult in real-world scenarios. Apart from high intrinsic similarity between the camouflaged objects and their background, the objects are usually diverse in scale, fuzzy in appearance, and even severely occluded. To deal with these problems, we propose a mixed-scale triplet network, \textbf{ZoomNet}, which mimics the behavior of humans when observing vague images, i.e., zooming in and out. Specifically, our ZoomNet employs the zoom strategy to learn the discriminative mixed-scale semantics by the designed scale integration unit and hierarchical mixed-scale unit, which fully explores imperceptible clues between the candidate objects and background surroundings. Moreover, considering the uncertainty and ambiguity derived from indistinguishable textures, we construct a simple yet effective regularization constraint, uncertainty-aware loss, to promote the model to accurately produce predictions with higher confidence in candidate regions. Without bells and whistles, our proposed highly task-friendly model consistently surpasses the existing 23 state-of-the-art methods on four public datasets. Besides, the superior performance over the recent cutting-edge models on the SOD task also verifies the effectiveness and generality of our model. The code will be available at \url{https://github.com/lartpang/ZoomNet}.

1. Introduction

Camouflaged object detection is difficult because objects blend into complex backgrounds and vary in scale, appearance, and occlusion. ZoomNet addresses these challenges with mixed-scale feature processing and uncertainty-aware optimization.

  • Camouflaged objects blend into their environments through altered appearance, coloration, or patterns, making them difficult to distinguish.
  • COD is more challenging than conventional salient object detection because targets can be tiny, background-like, or heavily obscured.
  • Existing deep learning methods still struggle with accurate and reliable detection because of visual insignificance, scale and appearance diversity, occlusion, and background distractors.
  • ZoomNet mimics human zooming by extracting scale-specific features with a triplet architecture, screening and aggregating them using SIUs, and enhancing mixed-scale features with HMUs.
  • The model uses UAL alongside BCE-based supervision to reduce uncertainty and interference from indistinguishable textures without adding parameters.
  • ZoomNet surpasses 23 state-of-the-art methods on four COD datasets under seven metrics and also generalizes effectively to SOD.

2. Related Work

Camouflaged object research builds on salient object detection but must handle objects that are visually inconspicuous and structurally ambiguous. ZoomNet differs by explicitly combining differentiated multi-scale cues and uncertainty-aware supervision.

  • Camouflaged object research concerns targets that evade detection through natural concealment, while SOD primarily addresses salient objects.
  • The method combines mixed-scale semantic features with BCE and uncertainty-aware loss to produce accurate and robust camouflaged-object segmentation, with transfer to SOD.
  • Scale Space Integration: Existing scale-space approaches combine inter-layer features to improve object localization and segmentation, but conventional inverted pyramids can lose texture and appearance details.
  • Scale Space Integration: For COD, prior approaches also overlook ambiguity in the structural information of the data, which creates a performance bottleneck.
  • Scale Space Integration: ZoomNet mimics zooming in and out to compare object-background relationships across scales while also exploring fine-grained feature scale space between channels.

3. Proposed Method

ZoomNet uses a triplet architecture to aggregate information across zoom scales, then refines mixed-scale and channel-specific semantics before producing a confidence map. Its uncertainty-aware optimization addresses ambiguity in camouflaged-object predictions.

  • Overall Architecture: ZoomNet customizes an image-pyramid-inspired triplet architecture with main and auxiliary scales to aggregate differentiated information for COD.The empirically selected scales are 1.0×, 1.5×, and 0.5×, processed by shared E-Net and C-Net feature encoders.
  • Hierarchical Mixed-scale Unit: Hierarchical mixed-scale units progressively integrate multi-level features top-down, increasing receptive-field range and diversifying feature representations.HMUs perform group-wise interaction followed by channel-wise modulation to strengthen scale-specific and subtle semantic representations.
  • Scale Integration Unit: The scale integration unit adaptively weights resized features from three scales using attention maps before integrating them.High-resolution features are downsampled with hybrid max- and average-pooling, while low-resolution features use bilinear interpolation.
  • Hierarchical Mixed-scale Unit: HMUs iteratively mix channel groups and use a learned modulation vector to refine concatenated channel features.The decoder uses five cascaded HMUs and stacked convolutions to produce logits, followed by a sigmoid-generated confidence map.
  • Loss Functions: The total training objective combines binary cross entropy with uncertainty-aware loss to reduce ambiguity and improve prediction confidence.Binary cross entropy supervises pixel predictions, while the uncertainty-aware term targets difficult, ambiguous samples.

4. Experiments

Experiments evaluate ZoomNet on four COD datasets, compare it with existing methods, and analyze its components, mixed-scale input scheme, and uncertainty-aware loss. Results report stronger quantitative and qualitative performance, while ablations attribute improvements to the proposed modules and UAL.

  • Datasets and evaluation: ZoomNet is evaluated on CAMO, CHAMELEON, COD10K, and NC4K, using seven common COD and SOD metrics.Training uses 3,040 COD10K images and 1,000 CAMO images, with remaining images used for testing.
  • Quantitative evaluation: ZoomNet consistently and significantly surpasses recent methods on all four datasets without post-processing tricks.It also improves over UJSC despite UJSC using extra SOD training data.
  • Quantitative evaluation: 19.3% average MAE improvement over C2F-Net is reported across four datasets, while average gains in Fωβ and Fβ are 4%.PR and Fβ curves also support the effectiveness of the proposed method.
  • Qualitative evaluation: ZoomNet produces clearer, more complete object regions and sharper contours across varying object sizes, occlusions, background interference, and indefinable boundaries.The qualitative comparisons cover large, middle, and small objects together with several difficult scene conditions.
  • Ablation studies: SIUs and HMUs each significantly improve the baseline, benefit one another, and help locate and distinguish camouflaged objects more accurately.Their cooperation with the loss functions enables capture of different object scales and predictions with higher contrast and consistency.
  • Ablation studies: The mixed-scale input scheme outperforms single-scale and simply mixed schemes on COD10K and CAMO, supporting its design for COD.The increasing cosine strategy performs best among the tested UAL weighting strategies, with smooth warm-up mitigating early interference with BCEL.
  • Ablation studies: UAL reduces ambiguous predictions: pixel values move toward two extremes, while feature maps become more discriminative and compact in camouflaged regions.Without UAL, middle-range histogram values and background interference correspond to blurred or uncertain predictions and false positives.
  • Discussion on SOD and COD: ZoomNet also performs outstandingly on SOD, which the authors attribute to shared demands for multi-scale, category-free discriminative features and reduced prediction ambiguity.The authors connect this generality to accurate localization, smooth segmentation, and UAL's mitigation of vagueness during detailed information recovery.

5. Conclusion

The conclusion presents ZoomNet as a zoom-inspired mixed-scale triplet network for camouflaged object detection. Its scale modules and uncertainty-aware loss support superior performance on both COD and SOD, while Figure 7 visualizes decoder feature-map effects.

  • 5. Conclusion: ZoomNet imitates human zooming in and out to analyze differentiated scene expressions at multiple scales.The conclusion frames this process as improving understanding and judgment of camouflaged objects.
  • 5. Conclusion: The scale merging layer filters and aggregates scale-specific features, while the hierarchical mixed-scale decoder mines mixed-scale semantics through grouping, mixing, and fusion.These stages are presented as the core feature-representation process.
  • 5. Conclusion: The uncertainty-aware loss penalizes prediction ambiguity, complementing the mixed-scale feature-processing components.The conclusion identifies it as the final optimization enhancement in the proposed pipeline.
  • 5. Conclusion: Extensive experiments verify ZoomNet's effectiveness on both COD and SOD, with superior performance to existing state-of-the-art methods.The appendix additionally reports SOD performance and provides further details.

A. Model Details

The model uses a ResNet50-based encoder, compressed multi-level features, and a decoder with hierarchical mixed-scale fusion. Its HMU iteratively mixes feature groups and applies channel-wise modulation while preserving feature dimensionality.

  • Encoder and compression: E-Net uses ResNet50 features before max-pooling and from layer1 through layer4, with channel counts increasing from 64 to 2048.The layers after layer4 are removed.
  • Encoder and compression: C-Net compresses encoder features using a simplified ASPP at layer4 and independent 3×3 CBR units elsewhere, producing 64 channels at every level.The ASPP uses five branches with dilation rates 1, 2, 5, 7, and 1, plus global-context pooling.
  • Decoder: The decoder progressively fuses up-sampled deeper features with shallower features, then applies a 1-channel convolution and sigmoid prediction layer.ZoomNet replaces the decoder fusion unit with HMU while retaining the other decoder layers.
  • Baseline: The ablation baseline combines E-Net, C-Net, and a basic convolutional decoder whose fusion units are 3×3 CBR blocks.The baseline is used to assess the proposed components.
  • Hierarchical mixed-scale unit: HMU iteratively mixes feature groups as an integrated multi-path kernel pyramid with partial parameter sharing, then modulates channels and branches using a learned vector.Its input and output features each have 64 channels, with 32 channels per group.

C.1. PR & Fβ curves of COD Methods

This section presents PR and Fβ curves across four COD datasets, alongside visualizations of decoder features, prediction histograms, and alternative UAL forms.

  • PR and Fβ curves: PR and Fβ curves compare different COD methods across four datasets, with ZoomNet represented by the red curve.The passage identifies the plotted methods and datasets but does not state a numerical ranking.
  • Feature visualization: The decoder’s intermediate feature maps are visualized to inspect representations from different stages.

C.7. Performance in More Complex Scenes

The evaluation targets complex COD scenes using diverse COD10K-TE scenarios and the small-scale CPD1K benchmark. Results on CPD1K-TE indicate strong performance in more complex scenarios.

  • Complex-scene benchmarks: COD10K-TE is described as a representative test dataset containing diverse scenarios and objects.
  • Complex-scene benchmarks: CPD1K is characterized as a complex small-scale dataset for testing adaptability to challenging scenarios.The compared methods are trained without CPD1K-TR.
  • Complex-scene benchmarks: The experiment reports superior performance for the proposed method on CPD1K-TE in more complex scenarios.Table 7 reports methods trained without CPD1K-TR.

D. Experiments on SOD

The authors evaluate ZoomNet on the salient object detection task to assess its generalization and the rationale of its structural design.

  • SOD evaluation: The proposed model is evaluated on SOD to examine generalization and the rationality of its structural design.

D.1. Datasets

The SOD experiments use five established datasets, with DUTS providing the training data and the remaining data used for inference.

  • D.1. Datasets: Five SOD datasets are used: DUT-OMRON, DUTS, ECSSD, HKU-IS, and Pascal-S.Their listed sizes are 5168, 10553 + 5017, 1000, 4447, and 850, respectively.
  • D.1. Datasets: Only the DUTS training set is used for model training.
  • D.1. Datasets: The remaining data are used for inference during testing.

D.2. Implementation Details

For SOD evaluation, ZoomNet is retrained under established training practices and compared against 22 methods using PR and Fβ curves and detailed results.

  • D.2. Implementation Details: ZoomNet is retrained on DUTS following the training strategies and techniques of prior methods.
  • D.2. Implementation Details: Figure 13 presents PR and Fβ curves for ZoomNet and recent state-of-the-art algorithms across five SOD datasets.
  • D.2. Implementation Details: The SOD comparison includes 22 existing methods, with results reported in Table 8 and Figure 13.
  • D.2. Implementation Details: ZoomNet outperforms all compared competitors on the more general binary segmentation task.

E. Limitations and Future Work

The authors identify inference efficiency and small-scale feature mining as remaining limitations, and propose simplifying the structure and improving contextual cue extraction.

  • E. Limitations and Future Work: The explicit scale-independent design adds inference cost, and ZoomNet is slightly slower than the fastest method, C2FNet.
  • E. Limitations and Future Work: The SIU still has room for improvement in mining effective clues from small-scale features.
  • E. Limitations and Future Work: Future work will simplify the inference structure and optimize contextual cue extraction from small-scale features.
Loading 2203.02688v1…