Object Detection
Region 기반 딥러닝 : R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN
Grid 기반 딥러닝 : YOLO, DetectNet
- YOLO는 이미지 내의 bounding box와 class probability를 single regression problem으로 간주하여, 이미지를 한 번 보는 것으로 오브젝트의 종류와 위치를 추측합니다
- 아래와 같이 single convolutional network를 통해 multiple bounding box에 대한 class probability를 계산하는 방식을 취합니다.
| 년도 | 알고리즘 | 링크 | 입력 | 출력 | 특징 |
|---|---|---|---|---|---|
| 2014 | R-CNN | 논문 | Image | Bounding boxes + labels for each object in the image. | AlexNet, 'Selective Search'사용 |
| 2015 | Fast R-CNN | 논문 | Images with region proposals. | Object classifications | Speeding up and Simplifying R-CNN, RoI Pooling |
| 2016 | Faster R-CNN | 논문 | CNN Feature Map. | A bounding box per anchor | MS, Region Proposal |
| YOLO | |||||
| SSD | Faster R-CNN + YOLO |
2017년 Mask R-CNN이 발표 되었지만 Segmentation 분야여서 포함 안함
참고 자료
- [추천] 딥러닝 참고 논문 요약 - Segmentation 등: sogangori 블로그
Recognition using Regions
J. Long, “Fully convolutional networks for semantic segmentation,” in IEEE CVPR 2015. :Github, 한글분석, Arxiv, ppt
K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image Recognition, CVPR 2016. (R-CNNs are based on ResNets)
| Method | VOC2007 | VOC2010 | VOC2012 | ILSVRC 2013 | MSCOCO 2015 | Speed |
|---|---|---|---|---|---|---|
| OverFeat | - | - | - | 24.3% | - | - |
| R-CNN (AlexNet) | 58.5% | 53.7% | 53.3% | 31.4% | - | - |
| R-CNN (VGG16) | 66.0% | - | - | - | - | - |
| SPP_net(ZF-5) | 54.2%(1-model), 60.9%(2-model) | - | - | 31.84%(1-model), 35.11%(6-model) | - | - |
| DeepID-Net | 64.1% | - | - | 50.3% | - | - |
| NoC | 73.3% | - | 68.8% | - | - | - |
| Fast-RCNN (VGG16) | 70.0% | 68.8% | 68.4% | - | 19.7%(@[0.5-0.95]), 35.9%(@0.5) | - |
| MR-CNN | 78.2% | - | 73.9% | - | - | - |
| Faster-RCNN (VGG16) | 78.8% | - | 75.9% | - | 21.9%(@[0.5-0.95]), 42.7%(@0.5) | 198ms |
| Faster-RCNN (ResNet-101) | 85.6% | - | 83.8% | - | 37.4%(@[0.5-0.95]), 59.0%(@0.5) | - |
| SSD300 (VGG16) | 72.1% | - | - | - | - | 58 fps |
| SSD500 (VGG16) | 75.1% | - | - | - | - | 23 fps |
| ION | 79.2% | - | 76.4% | - | - | - |
| CRAFT | 75.7% | - | 71.3% | 48.5% | - | - |
| OHEM | 78.9% | - | 76.3% | - | 25.5%(@[0.5-0.95]), 45.9%(@0.5) | - |
| R-FCN (ResNet-50) | 77.4% | - | - | - | - | 0.12sec(K40), 0.09sec(TitianX) |
| R-FCN (ResNet-101) | 79.5% | - | - | - | - | 0.17sec(K40), 0.12sec(TitianX) |
| R-FCN (ResNet-101),multi sc train | 83.6% | - | 82.0% | - | 31.5%(@[0.5-0.95]), 53.2%(@0.5) | - |
| PVANet 9.0 | 81.8% | - | 82.5% | - | - | 750ms(CPU), 46ms(TitianX) |
성능향상 기법
- Hard negative mining을 이용한 2~3% 성능 향상 방법들에 대한 서베이 논문