Auto Seg-Loss: Searching Metric Surrogates for Semantic Segmentation
This work addresses the problem of manual, expertise-intensive loss function design for semantic segmentation, offering an automated solution that improves model performance, though it is incremental in nature.
The paper tackles the misalignment between loss functions and evaluation metrics in semantic segmentation by automating the design of metric-specific surrogate losses through parameter search, resulting in consistent performance improvements on datasets like PASCAL VOC and Cityscapes.
Designing proper loss functions is essential in training deep networks. Especially in the field of semantic segmentation, various evaluation metrics have been proposed for diverse scenarios. Despite the success of the widely adopted cross-entropy loss and its variants, the mis-alignment between the loss functions and evaluation metrics degrades the network performance. Meanwhile, manually designing loss functions for each specific metric requires expertise and significant manpower. In this paper, we propose to automate the design of metric-specific loss functions by searching differentiable surrogate losses for each metric. We substitute the non-differentiable operations in the metrics with parameterized functions, and conduct parameter search to optimize the shape of loss surfaces. Two constraints are introduced to regularize the search space and make the search efficient. Extensive experiments on PASCAL VOC and Cityscapes demonstrate that the searched surrogate losses outperform the manually designed loss functions consistently. The searched losses can generalize well to other datasets and networks. Code shall be released.