Mining Cross-Image Semantics for Weakly Supervised Semantic Segmentation

Abstract

This paper studies the problem of learning semantic segmentation from image-level supervision only. Current popular solutions leverage object localization maps from classifiers as supervision signals, and struggle to make the localization maps capture more complete object content. Rather than previous efforts that primarily focus on intra-image information, we address the value of cross-image semantic relations for comprehensive object pattern mining. To achieve this, two neural coattentions are incorporated into the classifier to complimentarily capture cross-image semantic similarities and differences. In particular, given a pair of training images, one co-attention enforces the classifier to recognize the common semantics from co-attentive objects, while the other one, called contrastive co-attention, drives the classifier to identify the unshared semantics from the rest, uncommon objects. This helps the classifier discover more object patterns and better ground semantics in image regions. In addition to boosting object pattern learning, the co-attention can leverage context from other related images to improve localization map inference, hence eventually benefiting semantic segmentation learning. More essentially, our algorithm provides a unified framework that handles well different WSSS settings, i.e., learning WSSS with (1) precise image-level supervision only, (2) extra simple single-label data, and (3) extra noisy web data. It sets new state-of-the-arts on all these settings, demonstrating well its efficacy and generalizability.

Weakly supervised learning for image keypoint matching using graph convolutional networks

“…Datasets and Metrics: Following most of the prior works [3,4,5,6,7,10,11,12] we evaluate our method on the PAS-CAL VOC 2012 semantic segmentation benchmark [14]. It includes images with pixel-wise class labels, of which 1,464, 1,449, and 1,456 are used for training, validation, and test, respectively.…”

“…The backbone ResNet-101 is pre-trained on ImageNet. Baselines: We compare our approach with several stateof-the-art weakly-supervised image semantic segmentation methods [3,4,5,6,7,10,11,12] on training, validation, and test sets. The comparison with PSA [3] and IRNet [4] singles out the benefits of our feature propagation framework over label propagation since our GCN adopts the same affinity matrix as theirs for propagating features rather than activation Method mIoU (%) PSA [3] 59.7 SC-CAM [6] 63.4 SEAM [5] 63.6 IRNet [4] 66.5 SingleStage [7] 66.9 WSGCN-P 64.0 WSGCN-I 68.0 scores of CAMs.…”

“…Experimental results show that our complete pseudo labels have higher accuracy in Mean Intersection over Union (mIoU) than label propagation [3,4]. The net effect is that the semantic segmentation network trained with our complete pseudo labels outperforms the state-of-the-art baselines [5,6,7,10,11,12] on the PASCAL VOC 2012 dataset. In our implementation, the Affinity Network is from [3,4].…”

Weakly-Supervised Image Semantic Segmentation Using Graph Convolutional Networks

“…Experimental results show that our complete pseudo labels have higher accuracy in Mean Intersection over Union (mIoU) than label propagation [3,4]. The net effect is that the semantic segmentation network trained with our complete pseudo labels outperforms the state-of-the-art baselines [5,6,7,10,11,12] on the PASCAL VOC 2012 dataset.…”

Coffee Disease Visualization and Classification

“…Thus, weak supervision is more applicable and desirable for deep learning techniques [ 9 ]. This observation has sparked many weakly supervised learning studies [ 10 , 11 , 12 , 13 ], while other studies have attempted semi-supervised anomaly detection using autoencoders [ 14 ].…”