Command Palette
Search for a command to run...
Zhi Tian Chunhua Shen* Hao Chen Tong He

摘要
我们提出了一种全卷积单阶段目标检测器(FCOS),以像素级预测的方式解决目标检测问题,类似于语义分割。几乎所有的最先进目标检测器,如RetinaNet、SSD、YOLOv3和Faster R-CNN,都依赖于预定义的锚框。相比之下,我们提出的检测器FCOS既不使用锚框也不使用候选区域。通过消除预定义的锚框集合,FCOS完全避免了与锚框相关的复杂计算,例如在训练过程中计算重叠度。更重要的是,我们也避免了所有与锚框相关的超参数,这些超参数通常对最终的检测性能非常敏感。仅通过后处理中的非极大值抑制(NMS),使用ResNeXt-64x4d-101模型的FCOS在单模型和单尺度测试中实现了44.7%的平均精度(AP),超过了以往的单阶段检测器,并且具有更简单的优点。首次证明了一个更简单、更灵活的检测框架可以实现更高的检测精度。我们希望所提出的FCOS框架能够作为许多其他实例级任务的一个简单而强大的替代方案。代码可从以下链接获取:https://tinyurl.com/FCOSv1
代码仓库
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| 2d-object-detection-on-sardet-100k | FCOS | box mAP: 49.8 |
| object-detection-on-coco | FCOS (ResNeXt-32x8d-101-FPN) | AP50: 62.2 AP75: 46.1 APL: 52.6 APM: 45.6 APS: 26.0 Hardware Burden: Operations per network pass: box mAP: 42.7 |
| object-detection-on-coco | FCOS (ResNeXt-101-64x4d-FPN) | AP50: 62.8 AP75: 46.6 APL: 53.3 APM: 46.2 APS: 26.5 Hardware Burden: Operations per network pass: box mAP: 43.2 |
| object-detection-on-coco | FCOS (ResNeXt-64x4d-101-FPN 4 + improvements) | AP50: 64.1 AP75: 48.4 APL: 55.6 APM: 47.5 APS: 27.6 Hardware Burden: Operations per network pass: box mAP: 44.7 |
| object-detection-on-coco | FCOS (HRNet-W32-5l) | AP50: 60.4 AP75: 45.3 APL: 51.0 APM: 45.0 APS: 25.4 Hardware Burden: Operations per network pass: box mAP: 42.0 |
| object-detection-on-coco-minival | FCOS (ResNet-50-FPN + improvements) | AP50: 57.4 AP75: 41.4 APL: 49.8 APM: 42.5 APS: 22.3 box AP: 38.6 |
| object-detection-on-coco-o | FCOS (ResNet-50) | Average mAP: 16.7 Effective Robustness: 0.25 |
| pedestrian-detection-on-tju-ped-campus | FCOS | ALL (miss rate): 41.62 HO (miss rate): 81.28 R (miss rate): 31.89 R+HO (miss rate): 39.38 RS (miss rate): 69.04 |
| pedestrian-detection-on-tju-ped-traffic | FCOS | ALL (miss rate): 40.02 HO (miss rate): 63.73 R (miss rate): 24.35 R+HO (miss rate): 28.86 RS (miss rate): 37.40 |