Name
Abstract | ||
|---|---|---|
Large-scale variations on face detection make the performance of the model hard to improve. In this paper, we propose a novel balance training method for anchor-free face detection. Firstly, we propose a data augmentation called crop-mosaic to provide scale continuous, balance and controllable Gtboxes. Secondly, we design a density balance scale assignment strategy to divide the scalematching range for different scale faces. Thirdly, we introduce a scale normalization loss for comparing training performance with each levels. The proposed methods are implemented on Caffe. Extensive experiments on popular benchmarks WIDER FACE demonstrate that the performance of model training based on Mobilenet-v1 is close to the state-of-the-art face detectors. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1007/978-3-030-86608-2_39 | BIOMETRIC RECOGNITION (CCBR 2021) |
Keywords | DocType | Volume |
Face detection, Anchor-free, Data augmentation | Conference | 12878 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 5 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Chengpeng Wang | 1 | 0 | 0.34 |
| Chunyu Chen | 2 | 0 | 0.34 |
| Siyi Hou | 3 | 0 | 0.34 |
| Ying Cai | 4 | 0 | 0.34 |
| Menglong Yang | 5 | 109 | 10.49 |