PFDet: 2nd Place Solutions to Open Images Competition
?
7 likes?8,331 views
1. PFDet was the 2nd place solution to the Open Images 2018 object detection challenge, detecting 500 object classes in 1.7 million images using Faster R-CNN with Feature Pyramid Networks, SE-ResNeXt, expert models, and new losses. 2. Key aspects included co-occurrence loss to leverage unlabeled data, sigmoid loss instead of softmax, and cosine annealing which together improved average precision by over 20%. 3. The team used 512 V100 GPUs with multi-node batch normalization to train their best single model which achieved 33% of their full ensemble score.
PFDet: 2nd Place Solutions to Open Images Competition
- 1. Takuya Akiba*, Tommi Kerola*, Yusuke Niitani*, Toru Ogawa*, Shotaro Sano*, and Shuji Suzuki* *: Equal Contribution PFDet: 2nd Place Solution to Open Images Competition
- 2. : iwiwi Kaggle Tokyo Meetup #2
- 3. : Shotaro Sano (@g_votte)
- 5. ? Object Detection ? 500, 1.7M, Bounding Box 12M ? : 2018 7/3 ~ 8/30
- 8. ? ? : ? : C Bounding Boxes C Bounding Box ? ? ? Object Detection
- 9. Bounding Box : mean Average Precision ? Average Precision ? Average Precision C C Precision IoU = >= 0.5 Confidence TP/FP precision@k 0.9 TP 100% 0.8 TP 100% 0.75 FP 66% 0.7 TP 75% 0.67 FP 60% ´ ´ ´ Average Precision
- 10. Object Detection 2007-2012 2014- 2018
- 11. MS COCO Open Images # of classes 80 500 # of images 0.12M 1.7M Increase by more than x14 Object Detection
- 12. Faster RCNN
- 13. Bounding Box & Bounding Box
- 14. &
- 17. PFDet
- 18. Architecture Faster RCNN + Feature Pyramid Networks + SE-ResNeXt + Pyramid Scene Parsing + Context Head + Non-maximal Weighted Training Method SGD + Sigmoid Loss + Cosine Annealing + Co-occurrence Loss
- 19. Faster RCNN SE-ResNeXt Feature Pyramid Networks Expert Models Sigmoid Loss Cosine Annealing Co-occurrence Loss
- 20. SE-ResNeXt Feature Pyramid Networks Expert Models
- 21. SE-ResNeXt
- 24. Feature Pyramid Networks ? ResNet (& upsampling) feature map ? Region Proposal Network
- 25. Pressure Cooker: 17 images Person: 800k images 238 classes appear in <1000 images Image from https://storage.googleapis.com/openimages/web/factsfigures.html Expert Models
- 28. Suppression BB & class BB & class BB & class BB & class BB & class BB & class Bounding Box
- 30. Sigmoid Loss
- 31. Sigmoid Loss ? ? E.g., Bounding Box Football Ball
- 32. FC Softmax volleyball score football score ball score Cross Entropy football or ball volleyball score football score ball score FC Sigmoid Sigmoid Sigmoid Cross Entropy football and ballCross Entropy Cross Entropy SoftmaxLossSigmoidLoss
- 34. Cosine Annealing
- 37. ? ? Bounding Box Bounding Box
- 39. +22.7AP improvement on ＾Human Parts￣ on 47 part classes average +9.2AP improvement
- 40. Faster RCNN SE-ResNeXt Feature Pyramid Networks Expert Models Sigmoid Loss Cosine Annealing Co-occurrence Loss
- 42. Ensemble
- 44. MS COCO Open Images # of classes 80 500 # of images 0.12M 1.7M Increase by more than x14
- 45. HW MN-1 V100 (32GB) x512 Infiniband ¢ Best single model 33 ¢ 83% (v.s. 8 GPU) SW ChainerMN
- 46. Multi-node Batch Normalization ? BN ghost BN C mean, var ? segmentation, detection multi-node BN C BN mean, var ? multi-node BN
- 47. ? Faster R-CNN + FPN + etc ? Backbone: SE-ResNeXt (vs. NASNet, etc´´) ? LR Schedule: Cosine (vs. step, poly, etc´´) ? Suppression: NMW (vs. NMS, Soft NMS, etc´´) ? PSP, Context head ? : Co-occurrence loss ? : Multi-node BN, Linear LR scaling, warmup ? expert model
- 48. 1. tech report: Takuya Akiba, Tommi Kerola, Yusuke Niitani, Toru Ogawa, Shotaro Sano, Shuji Suzuki. PFDet: 2nd Place Solution to Open Images Challenge 2018 Object Detection Track. https://arxiv.org/abs/1809.00778 2. Co-occurrence loss tech report: Yusuke Niitani, Takuya Akiba, Tommi Kerola, Toru Ogawa, Shotaro Sano, Shuji Suzuki. Sampling Techniques for Large-Scale Object Detection from Sparsely Annotated Objects. https://arxiv.org/abs/1811.10862