[Paper Review] Boundary-aware Camouflaged Object Detection via Deformable Point Sampling

Minhyeok Lee, Suhwan Cho+4 more|arXiv (Cornell University)|Nov 22, 2022

Visual Attention and Saliency Detection4 citations

TL;DR

This paper proposes DPS-Net, a novel boundary-aware camouflaged object detection network that uses a Deformable Point Sampling transformer (DPS transformer) to enhance feature extraction by fusing local boundary details and global contextual information. The method achieves state-of-the-art performance on CHAMELEON, CAMO, and COD10K datasets with improved accuracy and inference efficiency.

ABSTRACT

The camouflaged object detection (COD) task aims to identify and segment objects that blend into the background due to their similar color or texture. Despite the inherent difficulties of the task, COD has gained considerable attention in several fields, such as medicine, life-saving, and anti-military fields. In this paper, we propose a novel solution called the Deformable Point Sampling network (DPS-Net) to address the challenges associated with COD. The proposed DPS-Net utilizes a Deformable Point Sampling transformer (DPS transformer) that can effectively capture sparse local boundary information of significant object boundaries in COD using a deformable point sampling method. Moreover, the DPS transformer demonstrates robust COD performance by extracting contextual features for target object localization through integrating rough global positional information of objects with boundary local information. We evaluate our method on three prominent datasets and achieve state-of-the-art performance. Our results demonstrate the effectiveness of the proposed method through comparative experiments.

Motivation & Objective

• Address the challenge of detecting camouflaged objects with minimal visual distinction from the background due to similar color and texture.
• Overcome limitations in existing models that struggle to capture fine-grained boundary details when object and background are visually similar.
• Improve model efficiency by reducing computational overhead while maintaining high detection accuracy through sparse, adaptive feature sampling.
• Enhance feature representation by integrating both local boundary information and global object location cues in a unified transformer-based framework.

Proposed method

• The DPS transformer employs a local extractor that uses deformable sampling points with learned offsets to focus on salient boundary regions, improving local feature extraction efficiency.
• A global extractor generates soft object and background region masks via global average pooling on masked feature regions, enabling robust global context learning.
• An aggregator fuses local boundary features and global contextual features to produce comprehensive representations that enhance object localization.
• The deformable point sampling mechanism dynamically selects informative points based on predicted offsets, reducing computation while increasing feature relevance.
• The boundary decoder and boundary feature module (BFM) are integrated to refine boundary predictions and guide feature fusion using predicted boundary maps.
• The model is implemented with both CNN-based (DPS-Net-R) and transformer-based (DPS-Net-P) encoders, enabling high-speed inference and strong performance.

Experimental results

Research questions

• RQ1 Can deformable point sampling improve the extraction of sparse, meaningful boundary features in camouflaged object detection?
• RQ2 How does combining local boundary features with global contextual information affect detection accuracy in low-contrast, camouflage-rich scenes?
• RQ3 To what extent does the proposed DPS transformer reduce computational cost while maintaining or improving performance compared to iterative or dense sampling methods?
• RQ4 Does the integration of a boundary decoder and BFM significantly enhance the model’s ability to distinguish subtle foreground-background boundaries?

Key findings

• The proposed DPS-Net achieves state-of-the-art performance on three benchmark datasets: CHAMELEON, CAMO, and COD10K, demonstrating superior detection accuracy.
• The model achieves 36.3 FPS inference speed with only 41.9 million parameters in the CNN-based version (DPS-Net-R), outperforming SegMaR [17] in both speed and parameter efficiency.
• The deformable point sampling method shows optimal performance at 3×3 reference points, with diminishing returns beyond that, confirming its efficiency and effectiveness.
• The ablation study confirms that using the boundary decoder with BFM improves mask prediction by effectively guiding feature fusion using boundary context.
• The DPS transformer significantly enhances feature representation, as shown by improved activation maps that better capture object boundaries compared to standard encoders.
• The DPS-Net-P variant with a PVTv2-B4 backbone achieves high performance comparable to SegMaR [17] but with faster inference speed and lower parameter count.