Name
Abstract | ||
|---|---|---|
This paperproposes a new end-to-end trainable matching network based on receptive field, RF-Net, to compute sparse correspondence between images. Building end-toend trainable matching framework is desirable and challenging. The very recent approach, LF-Net, successfully embeds the entire feature extraction pipeline into a jointly trainablepipeline, andproduces the state-of-the-artmatching results. This paper introduces two modifications to the structure of LF-Net. First, we propose to construct receptive feature maps, which lead to more effective keypoint detection. Second, we introduce a general loss function term, neighbor mask, to facilitate trainingpatch selection. This results in improved stability in descriptortraining. We trained RF-Net on the open dataset HPatches, and comparedit with other methods on multiple benchmark datasets. Experiments show that RF-Net outperforms existing state-of-the-artmethods. |
Year | DOI | Venue |
|---|---|---|
2019 | 10.1109/CVPR.2019.00832 | 2019 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2019) |
Field | DocType | Volume |
Receptive field,Computer vision,Pattern recognition,Computer science,Image matching,End-to-end principle,Artificial intelligence | Journal | abs/1906.00604 |
ISSN | Citations | PageRank |
1063-6919 | 3 | 0.35 |
References | Authors | |
0 | 8 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Xuelun Shen | 1 | 4 | 0.74 |
| Cheng Wang | 2 | 218 | 32.63 |
| Xin Li | 3 | 258 | 19.84 |
| Zenglei Yu | 4 | 3 | 0.35 |
| Jonathan Li | 5 | 798 | 119.18 |
| Chenglu Wen | 6 | 121 | 19.17 |
| Ming Cheng | 7 | 54 | 13.93 |
| Zijian He | 8 | 4 | 0.74 |