Network

level: CVPR CCF A author: Eddy IIg date: 2017 keyword:

• Optical Flow

---

Paper: FlowNET2.0不理解

Summary

1. 文章以实验主导,数据的顺序, 模型的叠加,各种算法结果

Note 去加强

• https://lmb.informatik.uni-freiburg.de/

Proble Statement

previous work:

• End-to-End optical flow estimation with CNNs was first proposed by [10]. Featuring a 3D CNN[30],unsupervised learning [1,33],carefullly designed rotationally invariant architectures, pyramidal approach based on the coarse-to-fine idea of variational methods[20]. No significantly outperform.
• An alternative approach to learning-based optical flow estimation using CNN to match image patches. reach good accuracy ,but require lots computing.
• CNN trained for per-pixel prediction tasks often produce noisy or blurry results. refinement can be obtained by stacking several CNNs on top of each other,leding improve results in human pose estimation,semantic instace segmentation.

Methods

• system overview:

【Qustion 1】Dataset Schedules ?

both the kind of data and the order in which it is presented during training.

Conclusion

• foces on training data and show that the schedule of presenting data during training is very import
• develop a stacked architecture that includes warping of the second image with intermediate optical flow
• elaborate on small displacements by introducing a subnetwork specializing on small motions.

---

Paper: Focal Loss

Summary

Research Objective

• Application Area:
• Purpose: One stage detectors applied over a regular dense sampling of object locations ,scales and aspectratios.

Proble Statement

• class imbalance is addressed in R-CNN like detectors by a two-stage cascade and sampling heuristics .The proposal stage(selective Search[35],EdgeBoxes[39],DeepMask[24,25],RPN[28]).The second classification stage (sampling heuristics,fixed foreground-to-background ratio,online hard example mining[31])
• training procedure is still dominated by easily classified background examples. some method( bootstrapping[33,29], hard example mining[37,8,31])

previous work:

• Classic Object Detectors:
  • sliding-window paradigm .applied convolutional neural networks to handwritten digit recognition.HOG[4] and integral channel features[5] for pedestrian detection.
• Two-stage Detectors: R-CNN , RPN , Faster R-CNN
• One-stage Detectors: OverFeat[30] the first, SDD[22,9] ,YOLO .
• Class Imbalance:
  • Training is inefficient as most location are easy negatives that contribute no useful learning signal
  • the easy negatives can overwhelm training and lead to degenerate models.
• Robust Estimation: focal loss designed to address class imbalance by down-weighting inliers,their contribution to total loss is small.

Methods

• system overview:

Conclusion

• discover that the extreme foreground-background class imbalance encountered during training of dense detectors is the central cause,we reshaping the standard cross entropy loss such that is down-weights the loss assigned to well-classified examples.
• training on a sparse set of hard examples and prevents the vast number os easy negatives from easy negatives from overwhelming the detector during training.
• design a simple-one-stage object detector called RetinaNet.

Notes 去加强了解

• https://github.com/facebookresearch/Detectron
• 4,5,8 传统方法
• 20 FPN method.