CamoTeacher: Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection

TL;DR

CamoTeacher tackles the high annotation cost in camouflaged object detection by introducing a semi-supervised framework that leverages Dual-Rotation Consistency Learning (DRCL). DRCL comprises Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL) to adaptively weight pseudo-labels based on rotation-view consistency, reducing both pixel-level and instance-level noise. Built on a Mean Teacher backbone, the method updates a teacher with EMA and optimizes a combined loss $L = L_s + \lambda_u L_u$, where $L_u$ includes DRCL terms. Across four COD benchmarks, CamoTeacher delivers strong improvements over baselines, even rivaling fully supervised models, and benefits further from additional unlabeled data and cross-model applicability to CNN and Transformer-based COD architectures.

Abstract

Existing camouflaged object detection~(COD) methods depend heavily on large-scale pixel-level annotations.However, acquiring such annotations is laborious due to the inherent camouflage characteristics of the objects.Semi-supervised learning offers a promising solution to this challenge.Yet, its application in COD is hindered by significant pseudo-label noise, both pixel-level and instance-level.We introduce CamoTeacher, a novel semi-supervised COD framework, utilizing Dual-Rotation Consistency Learning~(DRCL) to effectively address these noise issues.Specifically, DRCL minimizes pseudo-label noise by leveraging rotation views' consistency in pixel-level and instance-level.First, it employs Pixel-wise Consistency Learning~(PCL) to deal with pixel-level noise by reweighting the different parts within the pseudo-label.Second, Instance-wise Consistency Learning~(ICL) is used to adjust weights for pseudo-labels, which handles instance-level noise.Extensive experiments on four COD benchmark datasets demonstrate that the proposed CamoTeacher not only achieves state-of-the-art compared with semi-supervised learning methods, but also rivals established fully-supervised learning methods.Our code will be available soon.

Figures (5)

• Figure 1: (a) Illustration of pixel-level and instance-level noise. Pixel-level noise refers to varying degrees of noise in different parts within an instance. Instance-level noise pertains to varying degrees of noise among different instances. (b) Visualization of performance trends on COD10KCOD10K under different percentages of labeled data.
• Figure 2: (a) The real absolute error is obtained by subtracting the pseudo-label from the ground truth, reflecting the actual noisy level. On the other hand, pixel-wise inconsistency is derived from the subtraction of pseudo-labels from two rotated views. Visually, pixel-wise inconsistency and real absolute error appear very similar. (b) We define the area within 20 pixels from the edge as the boundary according to GT. We calculate mean pixel-wise inconsistency (MPI) and mean absolute error (MAE) in background, foreground, and boundary. The pseudo-labels are obtained by our proposed model on the test dataset COD10KCOD10K at a 10% semi-supervised setting.
• Figure 3: (a) The more similar the pseudo-labels are between two rotated views, the lower the noise level and the higher the quality of the pseudo-labels. For example, in the first row, the pseudo-labels are more similar, indicating a lower noise level compared to the pseudo-labels in the second row. (b) The positive correlation between instance-wise consistency and the SSIMSSIM of pseudo-labels and GT.
• Figure 4: The overall pipeline of our CamoTeacher. CamoTeacher consists of a teacher model and a student model. The teacher model generates pseudo-labels to optimize the student model, while the teacher model is updated through EMA mean-teacher from the student. To mitigate pseudo-labels' noise, we introduced Dual-Rotation Consistency Learning (DRCL), which involves Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). DRCL weights the loss function based on pixel-wise inconsistency and instance-wise consistency in pseudo-labels across different rotation views.
• Figure 5: Visualization of predictions in different proportions of labeled data.