PaperlayerView exact revision on arXiv

Source-linked AI summary

High-resolution Iterative Feedback Network for Camouflaged Object Detection

Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Ying Tai, Chengjie Wang, Ling Shao

arXiv:2203.11624v2cs.CV

TL;DR

Camouflaged object detection loses important edge and texture details when high-resolution inputs are degraded, motivating methods that preserve high-resolution information. HitNet iteratively feeds high-resolution features back into low-resolution representations and supervises each iteration, achieving improvements over 29 state-of-the-art methods on four challenging datasets while also enabling additional camouflaged training samples.

  • Problem

    Down-sampling high-resolution inputs loses edges and textures, causing blurry segmentation and motivating preservation of high-resolution information for COD.

  • Method

    HitNet uses cross-scale recursive feedback to refine low-resolution features with high-resolution information and applies an iterative feedback loss to constrain each iteration.

  • Results

    HitNet significantly outperforms 29 cutting-edge models across four challenging datasets; representative results include F w β of 0.798 on COD10K and mean MAE of 0.018 on CHAMELEON.

  • Takeaways & Limitations

    Cross-domain learning converts salient objects into camouflaged objects, providing additional COD training data without increasing inference-stage parameters or computation.

Abstract

from arXiv · show

Spotting camouflaged objects that are visually assimilated into the background is tricky for both object detection algorithms and humans who are usually confused or cheated by the perfectly intrinsic similarities between the foreground objects and the background surroundings. To tackle this challenge, we aim to extract the high-resolution texture details to avoid the detail degradation that causes blurred vision in edges and boundaries. We introduce a novel HitNet to refine the low-resolution representations by high-resolution features in an iterative feedback manner, essentially a global loop-based connection among the multi-scale resolutions. In addition, an iterative feedback loss is proposed to impose more constraints on each feedback connection. Extensive experiments on four challenging datasets demonstrate that our \ourmodel~breaks the performance bottleneck and achieves significant improvements compared with 29 state-of-the-art methods. To address the data scarcity in camouflaged scenarios, we provide an application example by employing cross-domain learning to extract the features that can reflect the camouflaged object properties and embed the features into salient objects, thereby generating more camouflaged training samples from the diverse salient object datasets The code will be available at https://github.com/HUuxiaobin/HitNet.

I. INTRODUCTION

Camouflaged object detection must distinguish foregrounds from highly similar surroundings while preserving fine boundary and texture details lost during down-sampling. HitNet addresses this with high-resolution iterative feedback, achieves strong benchmark results, and supports synthetic camouflaged-data generation.

  • Camouflaged object detection identifies hidden objects or animals that blend with their surroundings and supports applications including medical segmentation, artistic creation, species discovery, and crack inspection.
  • Down-sampling discards high-resolution edges and textures, producing blurry vision and less detailed boundaries than high-resolution inputs.High-resolution inputs produce more detailed predictions, including cat beards, than low-resolution inputs evaluated with the same model.
  • HitNet refines low-resolution features with high-resolution information through adaptive iterative feedback across multi-scale representations.Its components include Transformer-based Feature Extraction, multi-resolution iterative refinement, and iteration feature feedback.
  • Cross-domain learning converts salient objects into camouflaged objects to provide additional COD training data without increasing inference-stage parameters or computation.
  • HitNet imposes an iterative feedback loss that constrains each feedback connection by penalizing outputs from each iteration.
  • 0.798 weighted F-measure is achieved on COD10K, 16.5% higher than the second-best LSR, while CHAMELEON mean MAE is 0.018, 40.0% better than the second-best SINet-v2.

II. RELATED WORK

Related work progresses from handcrafted features to deep models that use contextual or boundary information, but the paper emphasizes preserving high-resolution information through HitNet’s feedback design. The framework combines multi-scale Transformer features, iterative refinement, and iteration-level supervision.

  • II. RELATED WORK: Early COD methods use handcrafted color, intensity, texture, convexity, or motion-boundary features but are less robust in complex scenarios.
  • II. RELATED WORK: HitNet uses a Transformer-based backbone for multi-scale feature extraction and cross-resolution iterative feedback to refine low-resolution features with high-resolution information.
  • II. RELATED WORK: The framework supervises each feedback iteration with a corresponding loss to constrain the feedback connections and reduce feature corruption.

A. Transformer-based Feature Extraction

HitNet extracts multi-scale features with a Pyramid Vision Transformer and repeatedly transports information across resolutions to refine low-resolution representations. Iteration-level segmentation supervision and high-resolution feedback preserve cues important for detailed COD boundaries.

  • A. Transformer-based Feature Extraction: Pyramid Vision Transformer extracts multi-scale features while reducing memory costs through progressive shrinking and spatial reduction attention.
  • A. Transformer-based Feature Extraction: The backbone produces four feature scales at 1/32, 1/16, 1/8, and 1/4 of the input resolution for subsequent processing.
  • A. Transformer-based Feature Extraction: A basic block processes each scale using two stacked 3 × 3 convolutions and channel attention.
  • A. Transformer-based Feature Extraction: Feedback features from previous iterations are up-sampled, concatenated with scale features, and compressed to match the target scale.
  • A. Transformer-based Feature Extraction: Feedback flows across scales and iterations, with explicit imports into the top and third-top scales while the second-top scale receives implicit feedback from the top scale.The setting reduces computational cost while maintaining good performance according to the paper.
  • A. Transformer-based Feature Extraction: Each iteration outputs an intermediate high-resolution segmentation map supervised by segmentation loss, allowing feedback features to learn high-resolution cues.
  • A. Transformer-based Feature Extraction: High-resolution feedback features merge with input features and use feedback fusion to exploit high-resolution data flow in the multi-scale structure.

C. Iteration Feature Feedback

The iteration feature feedback strategy supervises feedback features and predictions at each iteration to control recurrent feature flow and emphasize deeper iterations.

  • Iteration feature feedback ties each feedback feature to the segmentation ground truth to control feature flow and reduce recurrent-path corruption.
  • The basic loss combines weighted intersection-over-union and weighted binary cross entropy losses.
  • HitNet computes the high-resolution prediction loss at every iteration and applies an iteration-weight scheme to penalize each output.
  • The weighting scheme assigns higher weights to deeper iterations, focusing optimization on later features.
  • Non-local graph fusion integrates features from the previous module through a graph fusion module.

A. Experimental Settings

Experiments use four widely used camouflaged-object detection datasets, standard evaluation metrics, and a high-resolution 704×704 implementation setup.

  • Datasets: The experiments are based on four widely used COD datasets, including CHAMELEON and CAMO, with CAMO containing 2,500 images split into 2,000 training and 500 testing images.CHAMELEON contains 76 high-resolution images with manually annotated object-level ground-truth masks.
  • Implementation Details: Table I compares HitNet with 29 state-of-the-art methods across four benchmark datasets, highlighting the best and second-best results.
  • Datasets: The combined CAMO and COD10K training set is used for training, while the other datasets serve as test sets.
  • Metrics: Performance is comprehensively evaluated using MAE, mean E-measure Eφ, S-measure Sα, and weighted F-measure Fwβ.
  • Implementation Details: Input images are resized to 704×704 to avoid high-resolution information loss to some extent, without data augmentation.The model uses PVT-V2 initialization for transformer-based feature extraction, while remaining modules are randomly initialized.

B. Quantitative Evaluation

HitNet outperforms second-best methods across CHAMELEON, CAMO, NC4K, and COD10K, with improvements attributed to high-resolution information and iterative feedback.

  • CHAMELEON: 40.0% lower MAE error and 7.6% higher Fwβ than second-best models on CHAMELEON establish a performance milestone for HitNet.
  • CAMO: 18.6% lower MAE error and 7.8% higher Fwβ than the second-best model are achieved on CAMO.
  • NC4K: 18.7% lower MAE error and 5.9% higher Fwβ than second-best models are achieved on NC4K, demonstrating HitNet’s generalization ability.
  • COD10K: 31.4% lower MAE error and 16.5% higher Fwβ than second-best models are achieved on the challenging COD10K test set.
  • Overall Findings: HitNet’s superiority across four benchmarks is attributed mainly to exploiting high-resolution information and mitigating feature-level high-resolution degradation through iterative feedback.

C. Qualitative Evaluation

HitNet is evaluated qualitatively and through ablations, with emphasis on preserving fine boundaries and assessing iterative feedback, backbone, resolution, and computational choices.

  • Qualitative results: HitNet segments clear leaves, thorns, occluded objects, multiply camouflaged objects, and thin insect edges where other models produce blurred or incomplete details.These examples cover complex topology, dense thorns, occlusion, and globally thin edges.
  • Iterative feedback: The iterative feedback mechanism progressively improves representations across iterations, including sharper edges in later iterations.The default setting uses T = 4 to balance inference time and performance.
  • Input resolution: With the same degraded 352×352 inputs, HitNet reduces MAE error by 21.6% relative to the second-best SINet-V2.This result indicates that the model remains strongest under the controlled low-resolution comparison described.

V. APPLICATION

The application uses cross-domain learning to convert salient objects into camouflaged objects, with a contrastive index helping select suitable converted samples.

  • Application: Cross-domain learning converts salient objects into camouflaged objects by learning camouflaged features and embedding them into salient objects.The pipeline distinguishes salient and camouflaged domains using domain-specific discriminators.
  • Sample selection: A contrastive index evaluates camouflage level from pixel-level similarity between an object and its surroundings.The index is used to discard hard cases whose intrinsic salient appearance remains unchanged.

A. Cross-domain Learning

The cross-domain learning strategy uses cycle-consistent translation between salient and camouflaged domains and filters generated samples with a contrastive index. It improves training outcomes compared with directly adding salient data.

  • Cross-domain translation: Cycle-consistent learning translates salient samples toward the camouflaged domain and camouflaged samples toward the salient domain using generators and discriminators.G maps salient samples toward camouflaged images, while F performs the reverse translation; discriminators distinguish translated from real samples.
  • Contrastive selection: The contrastive index measures camouflage level using pixel intensities within one σ rule to reduce the effect of extreme values.Its defined quantities include pixel intensity Pi, image mean Pm, standard deviation Pstd, and pixel index i.
  • Contrastive selection: The index threshold is set to Isc = 20, allowing failure cases with highly salient converted objects to be discarded.An abandoned car is cited as a high-salient example rejected by the index.
  • Evaluation: Directly adding salient data worsens Sα from 0.868 to 0.844 and Eφ from 0.932 to 0.911, whereas CDL improves F wβ from 0.798 to 0.806 and reduces MAE error by 4.2%.The comparison supports using converted camouflaged samples rather than salient data alone for COD training.

VI. FAILURE CASE

HitNet still struggles with highly complex camouflaged structures and very small objects. Increasing iterative refinement beyond four iterations yields only slight performance gains while increasing runtime.

  • HitNet fails on challenging examples such as caterpillars with complicated topological structures and dense edges or details.
  • The method remains limited on objects with global long-range edges or multiple camouflaged objects, including very small objects that may be neglected.
  • The iterative feedback mechanism is a bottleneck because it self-corrects low-resolution features using high-resolution information for dense edges and small objects.
  • Beyond four iterations, performance improves only slightly, whereas computation time continues to increase.
  • Despite these failure cases, HitNet significantly outperforms 29 cutting-edge models across four challenging datasets.
Loading 2203.11624v2…