ParseNet: Looking Wider to See Better Wei Liu, Andrew Rabinovich, Alexander C. Berg This paper presents ParseNet, a simple and effective convolutional neural network for semantic segmentation. The key idea is to add global context to fully convolutional networks (FCNs) by using the average feature of a layer to augment features at each location. This approach significantly improves the performance of baseline networks, such as FCN, and achieves state-of-the-art results on SiftFlow and PASCAL-Context with minimal additional computational cost. On PASCAL VOC 2012, ParseNet achieves near state-of-the-art performance with a simple approach. Semantic segmentation involves per-pixel labeling of image content. FCN is a successful technique for semantic segmentation, but it ignores global context. To address this, ParseNet integrates global context into the FCN framework by pooling the feature map of a layer to create a context vector, which is then appended to the features sent to the next layer. This technique allows for end-to-end training and improves performance without significant computational overhead. The paper also discusses the importance of normalization in combining features from different layers. By normalizing each feature and learning the scaling parameters, the training becomes more stable and the performance improves. The approach is validated on three benchmark datasets: VOC2012, PASCAL-Context, and SiftFlow. Results show that adding global context improves performance, and that normalization is crucial for effective feature combination. ParseNet is compared with other methods, including DeepLab and DeepLab-LargeFOV, and is shown to achieve better performance with a simpler structure. The paper also discusses the importance of training and validation procedures, and the benefits of using normalization and learning weights when combining features from multiple layers. Overall, ParseNet demonstrates that adding global context to FCNs can significantly improve performance, and that normalization is essential for effective feature combination. The approach is simple, robust, and effective, achieving state-of-the-art results on several benchmark datasets.ParseNet: Looking Wider to See Better Wei Liu, Andrew Rabinovich, Alexander C. Berg This paper presents ParseNet, a simple and effective convolutional neural network for semantic segmentation. The key idea is to add global context to fully convolutional networks (FCNs) by using the average feature of a layer to augment features at each location. This approach significantly improves the performance of baseline networks, such as FCN, and achieves state-of-the-art results on SiftFlow and PASCAL-Context with minimal additional computational cost. On PASCAL VOC 2012, ParseNet achieves near state-of-the-art performance with a simple approach. Semantic segmentation involves per-pixel labeling of image content. FCN is a successful technique for semantic segmentation, but it ignores global context. To address this, ParseNet integrates global context into the FCN framework by pooling the feature map of a layer to create a context vector, which is then appended to the features sent to the next layer. This technique allows for end-to-end training and improves performance without significant computational overhead. The paper also discusses the importance of normalization in combining features from different layers. By normalizing each feature and learning the scaling parameters, the training becomes more stable and the performance improves. The approach is validated on three benchmark datasets: VOC2012, PASCAL-Context, and SiftFlow. Results show that adding global context improves performance, and that normalization is crucial for effective feature combination. ParseNet is compared with other methods, including DeepLab and DeepLab-LargeFOV, and is shown to achieve better performance with a simpler structure. The paper also discusses the importance of training and validation procedures, and the benefits of using normalization and learning weights when combining features from multiple layers. Overall, ParseNet demonstrates that adding global context to FCNs can significantly improve performance, and that normalization is essential for effective feature combination. The approach is simple, robust, and effective, achieving state-of-the-art results on several benchmark datasets.