Finding Camouflaged Objects Along the Camouflage Mechanisms

Common mechanisms for achieving object camouflage include reducing differences and increasing distractions. Such camouflage mechanisms hinder the object detectors to accurately distinguish the camouflaged objects from their surroundings. Considering that, we reexamine the camouflaged object detection (COD) task from the perspective of camouflage mechanisms and make the first attempt to discover the target objects in a de-camouflaging manner. We argue that this process can not only lead to a better understanding of camouflage, but also provide a new perspective for detecting camouflaged objects. For that, we first analyze some existing camouflage mechanisms together with their induced problems. Afterwards, considering the inner relationships between SOD and COD, we resort to the SOD task to synergistically achieve de-camouflaging for COD. Specifically, we incorporate the SOD task into the COD model and present a multi-task learning framework for COD, which models the intrinsic relationships between the two tasks from different perspectives, i.e., task-conflicting attribute and task-consistent attribute, to destroy the camouflage conditions for highlighting those inconspicuous yet valuable cues of camouflaged objects. In more detail, modeling the task-conflicting attribute is to well identify camouflaged objects by alleviating such interfering information from salient ones, and is achieved by a Gate Classification (GC) strategy and a Region Distraction Module (RDM). While, modeling the task-consistent attribute, which is achieved by an adversarial learning (AL) scheme and a Boundary Injection Module (BIM), is intended to enhance the boundary differences between the camouflaged objects and their backgrounds for fully segmenting the camouflaged objects. Extensive results demonstrate the superiorities of our proposed model over existing ones in camouflaged object detection.