From image-level to pixel-level labeling with Convolutional Networks

Abstract

We are interested in inferring object segmentation by leveraging only object class information, and by considering only minimal priors on the object segmentation task. This problem could be viewed as a kind of weakly supervised segmentation task, and naturally fits the Multiple Instance Learning (MIL) framework: every training image is known to have (or not) at least one pixel corresponding to the image class label, and the segmentation task can be rewritten as inferring the pixels belonging to the class of the object (given one image, and its object class). We propose a Convolutional Neural Network-based model, which is constrained during training to put more weight on pixels which are important for classifying the image. We show that at test time, the model has learned to discriminate the right pixels well enough, such that it performs very well on an existing segmentation benchmark, by adding only few smoothing priors. Our system is trained using a subset of the Imagenet dataset and the segmentation experiments are performed on the challenging Pascal VOC dataset (with no fine-tuning of the model on Pascal VOC). Our model beats the state of the art results in weakly supervised object segmentation task by a large margin. We also compare the performance of our model with state of the art fully-supervised segmentation approaches.

WILDCAT: Weakly Supervised Learning of Deep ConvNets for Image Classification, Pointwise Localization and Segmentation

“…We can see that WILDCAT without CRF outperforms existing weakly supervised models by a large margin. We note a large gain with respect to MIL models based on (soft-)max pooling [49,50], which validates the relevance of our pooling for segmentation. The improvement between WILD-CAT with CRF and the best model is 7.1 pt.…”

“…Mean IoU MIL-FCN [49] 24.9 MIL-Base+ILP+SP-sppxl [50] 36.6 EM-Adapt +FC-CRF [45] 33.8 CCNN + FC-CRF [48] 35.3 WILDCAT 39.2 WILDCAT + FC-CRF 43.7 With a quite more complex strategy, the very recent paper [31] presents impressive results (50.7 MIoU). The training scheme in [31] incorporates different terms, which are specifically tailored to segmentation: one enforces the segmentation mask to match low-level image boundaries, another one incorporates prior knowledge to support predicted classes to occupy a certain image proportion.…”

“…Another important WSL application is segmentation. Many methods are based on MIL framework: MIL-FCN [49] extends MIL to multi-class segmentation, MIL-Base [50] introduces a soft extension of MIL, EM-Adapt [45] includes an adaptive bias into the MIL framework, and Constrained CNN (CCNN) [48] uses a loss function optimized for any set of linear constraints on the output space of a CNN.…”

Exploiting Saliency for Object Segmentation from Image Level Labels

“…Another approach considers modifying the classifier training procedure so as to have it generate object masks as byproduct of a forward-pass. This can be achieved by adding a global a max-pooling [33] or mean-pooling layer [54] in the last stages of the classifier. In this work we provide an empirical comparison of existing seeders, and explore variants of the mean-pooling approach [54] …”

“…Strongly co-occurring categories (such as train and rails, sculls and oars, snowbikes and snow) cannot be separated without additional information. Because additional information is needed to solve the task, previous work have explored different avenues, including class-specific size priors [31], crawling additional images [33,46], or requesting corrections from a human judge [17,37].…”