The Lov´asz-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks

By Maxim Berman, Amal Rannen Triki, Matthew B. Blaschko

Subfields: Computer Vision and Pattern Recognition, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Optimization & Training

Citations in the corpus

Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression (2019, methodology): “Optimizing IoU using an approximate or a surrogate function: In the semantic segmentation task, there have been some efforts to optimize IoU using either an approximate function [18] or a surrogate loss [17].”
Self-Correction for Human Parsing (2019, methodology): “To tackle this issue, following by [1], we additionally introduce a tractable surrogate loss function for optimizing the mIoU directly.”
CycleSegNet: Object Co-Segmentation With Cycle Refinement and Region Correspondence (2021, unknown): “Please refer to [3] for more details about Lov ´ a sz-Softmax loss function.”
SalsaNext: Fast, Uncertainty-Aware Semantic Segmentation of LiDAR Point Clouds (2020, unknown): “In [2], the authors adopt this metric with the help of the Lov ´ asz extension for submodular functions.”
Combo Loss: Handling Input and Output Imbalance in Multi-Organ Segmentation (2018, methodology): “In this paper, we make the following contributions: a) We introduce a curriculum learning based loss function to [7] Weighted cross entropy 3D U-Net [8] Weighted cross entropy 3D V-net [9] Dice Brosch et. al [18] Sensitivity + speciﬁcity Sudre et. al [13] Generalized/weighted Dice Berman et. al [19]”
Semi-Supervised Segmentation of Salt Bodies in Seismic Images using an Ensemble of Convolutional Neural Networks (2019, methodology): “After that, we minimize Lovasz loss function [4] for 150 epochs, which allows a direct optimization of the IoU metric.”
Breast Tumor Cellularity Assessment using Deep Neural Networks (2019, methodology): “Rakhlin et al. [33] used the Resnet-34 network [34] as the encoder and the Lov ´ asz-Softmax loss function [35] along with Stochastic Weight Averaging (SWA) [36] for training.”
Salient Object Detection in the Deep Learning Era: An In-Depth Survey (2019, background): “Another recent work [138] proposes to directly optimize the mean intersection-over-union loss, which brings impact to semantic segmentation as well as its binary case, i.e . foreground-background segmentation.”
STEm-Seg: Spatio-temporal Embeddings for Instance Segmentation in Videos (2020, methodology): “However, we employ the Lov`asz hinge loss [3,4,59,104] (details in supplementary material), which is a differentiable, convex surrogate of the Jaccard index that directly optimizes the IoU.”
Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation (2020, methodology): “For the voxel-wise loss ( L voxel ), we follow the existing methods [10, 16] and use the weighted cross-entropy loss and lovasz-softmax [4] loss to maximize the point accuracy and the intersection-over-union score, respectively.”
ChangeMamba: Remote Sensing Change Detection With Spatiotemporal State Space Model (2024, background): “In addition, the Lovasz-softmax loss [54] is introduced to alleviate the problem of imbalance in the number of samples between changed and unchanged pixels.”
Optimization for Medical Image Segmentation: Theory and Practice When Evaluating With Dice Score or Jaccard Index (2020, methodology): “The metric-sensitive losses are typically expected to help most with segmentation of small-sized objects because they are invariant to scale [12].”
Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction (2023, methodology): “We employ both cross entropy (CE.) loss and lovasz-softmax loss [3] for LiDAR segmentation.”
SuperFusion: A Versatile Image Registration and Fusion Network with Semantic Awareness (2022, methodology): “Thus, we introduce the Lovasz-Softmax loss [74] to calculate the error between the predicted results and ground truth.”
Optimizing the Dice Score and Jaccard Index for Medical Image Segmentation: Theory and Practice (2019, background): “ISLES 2017 omitted for lack of statistical relevance given its lower number of samples. their invariance to scale, which cross-entropy does not have [5].”