Background Noise Reduction of Attention Map for Weakly Supervised Semantic Segmentation

TL;DR

This paper tackles background noise in attention weights within weakly supervised semantic segmentation by integrating attention-map-enhanced CAMs into the training loss. Utilizing a Conformer backbone, it refines CAMs with global attention, introduces noise-augmented CAMs during training, and optimizes with a combination of $L_{M^{**}}$ and $L_{cls}$ to suppress background regions. The approach yields improved segmentation results on PASCAL VOC 2012 and MS COCO 2014, surpassing the TransCAM baseline in both seed generation and final pseudo-label quality. The work demonstrates that incorporating attention-guided CAMs into the learning objective can meaningfully reduce background noise, enhancing pseudo-label accuracy and downstream segmentation performance.

Abstract

In weakly-supervised semantic segmentation (WSSS) using only image-level class labels, a problem with CNN-based Class Activation Maps (CAM) is that they tend to activate the most discriminative local regions of objects. On the other hand, methods based on Transformers learn global features but suffer from the issue of background noise contamination. This paper focuses on addressing the issue of background noise in attention weights within the existing WSSS method based on Conformer, known as TransCAM. The proposed method successfully reduces background noise, leading to improved accuracy of pseudo labels. Experimental results demonstrate that our model achieves segmentation performance of 70.5% on the PASCAL VOC 2012 validation data, 71.1% on the test data, and 45.9% on MS COCO 2014 data, outperforming TransCAM in terms of segmentation performance.

Figures (5)

• Figure 1: Comparison of CAMs generated by TransCAM and our proposed method. (a)Input image, (b)Ground Truth, (c)TransCAM, (d)CAMs by our method.
• Figure 2: Overview of the proposed method. Create TransCAM from input images. Add noise to TransCAM and input it to the loss function. Train TransCAM using image-level lable. Attention map is explicitly incorporated into the training process to suppress the background noise derived from the Attention map. Pseudo-labels are created from the initial seed obtained from TransCAM. The trained TransCAM has reduced background noise.
• Figure 3: Qualitative segmentation results on the PASCAL VOC 2012 validation set. (a) Original images; (b) Ground truth; (c)Prediction of Ours.
• Figure 4: mIoU variation according to Hard thresholds
• Figure 5: Comparisons of Attention map. (a) Original image; (b)Attention map generated by baseline; (c) Attention map generated by our method.