Improvement in design and training of feature pyramid network for contour refinement

Abstract

• A simplified feature pyramid architecture, called Edge Pyramid Network, applied for the first time to edge detection. • A recursive fine-tuning strategy introducing a new weighted negative loss to detect accurate and thin edges. • State-of-the-art results on the benchmark datasets for semantic and category-agnostic edge detection. Semantic edge detection is a challenging problem which aims to produce multiple edge maps corresponding to different object categories. The design and the training of the related network are the two most powerful levers to optimize its accuracy. To this end, we first refine existing semantic edge detection architectures with low and high level features inside our Edge Pyramid Network, which already outperforms state-of-the-art networks on popular benchmark datasets. Then, we apply an ingenious fine-tuning strategy, which incrementally improves edge detection accuracy through a few alternated training cycles, using various loss functions to perform thick and thin edge detection. We observe that applying separately any of these refinements outperforms such networks as SEAL and STEAL for category-aware edge detection. More remarkable is the fact that these two refinements provide accumulated benefits on the edge detection performance (+0.85% and +1.78% on MF-scores compared to STEAL on SBD and Cityscapes datasets respectively). This improvement is visible in the edge maps in which EPN architecture results in smoother edges and the fine-tuning strategy is responsible for thinner edges. Furthermore, we show that EPN architecture achieves competitive performance (ODS F-measure of 0.828) against the state-of-the-art category-agnostic edge detection networks on the BSDS500 dataset.