[Paper Review] Camouflaged Object Segmentation with Distraction Mining

Haiyang Mei, Ge-Peng Ji+4 more|arXiv (Cornell University)|Apr 21, 2021

Visual Attention and Saliency DetectionComputer Science67 references32 citations

TL;DR

PFNet introduces a bio-inspired positioning and focus framework for camouflaged object segmentation, employing distraction mining to suppress false positives and recover false negatives, achieving state-of-the-art results at real-time speed.

ABSTRACT

Camouflaged object segmentation (COS) aims to identify objects that are "perfectly" assimilate into their surroundings, which has a wide range of valuable applications. The key challenge of COS is that there exist high intrinsic similarities between the candidate objects and noise background. In this paper, we strive to embrace challenges towards effective and efficient COS. To this end, we develop a bio-inspired framework, termed Positioning and Focus Network (PFNet), which mimics the process of predation in nature. Specifically, our PFNet contains two key modules, i.e., the positioning module (PM) and the focus module (FM). The PM is designed to mimic the detection process in predation for positioning the potential target objects from a global perspective and the FM is then used to perform the identification process in predation for progressively refining the coarse prediction via focusing on the ambiguous regions. Notably, in the FM, we develop a novel distraction mining strategy for distraction discovery and removal, to benefit the performance of estimation. Extensive experiments demonstrate that our PFNet runs in real-time (72 FPS) and significantly outperforms 18 cutting-edge models on three challenging datasets under four standard metrics.

Motivation & Objective

• Motivate COS by mimicking predation detection and identification stages.
• Develop a two-module network (Positioning Module and Focus Module) to locate and refine camouflaged targets.
• Introduce a distraction mining strategy to discover and remove false distractions.
• Achieve real-time COS with superior accuracy on multiple benchmarks.
• Demonstrate ablations showing the contributions of PM and FM components.

Proposed method

• A ResNet-50 backbone extracts multi-level features from RGB images.
• Positioning Module uses channel and spatial non-local attention to generate an initial target location map.
• Focus Module performs distraction mining with foreground- and background-attentive features and context exploration blocks to discover false positives/negatives.
• Distraction discovery uses four-branch context exploration blocks with multi-scale dilated convolutions to capture wide contexts.
• Distraction removal combines distraction-removed features via subtraction/addition and refinement through learnable scalars.
• Loss combines PM losses (BCE + IoU) and FM losses (weighted BCE + weighted IoU) with a progressive multi-level weighting scheme.
• Training uses SGD, ImageNet-pretrained backbone, 416x416 inputs, no post-processing like CRF, and 72 FPS inference.

Experimental results

Research questions

• RQ1 Can a bio-inspired positioning-and-focusing framework improve camouflaged object segmentation over state-of-the-art methods?
• RQ2 How does distraction mining affect the handling of false positives and false negatives in COS?
• RQ3 What is the contribution of the PM and FM components to overall performance?
• RQ4 Is the PFNet capable of real-time inference on standard COS benchmarks?
• RQ5 How do ablations of attention blocks and distraction streams impact performance?

Key findings

• PFNet achieves state-of-the-art results on three COS benchmarks: CHAMELEON, CAMO, and COD10K using four metrics (S_alpha, E_ad, F_w, M).
• PFNet significantly outperforms SINet and other baselines, e.g., on CHAMELEON: S_alpha 0.882, E_ad 0.942, F_w 0.810, M 0.033; on CAMO: S_alpha 0.782, E_ad 0.852, F_w 0.695, M 0.085; on COD10K: S_alpha 0.800, E_ad 0.868, F_w 0.660, M 0.040.
• Inference runs at 72 FPS, faster than SINet (72 vs 51 FPS).
• Ablation studies show that both the Positioning Module and Focus Module (including distraction mining components) substantially improve performance.
• Distraction mining (false-positive and false-negative) with contextual exploration blocks yields clearer object delineation and better handling of camouflaged structures.
• Qualitative results demonstrate improved segmentation of small, large, and multi-object camouflage scenarios.