计算机视觉中常见问题及论文中的解决方案［不断更新，欢迎补充］

 

The order of solusions of each issue follows the timeline. You can see there is a trend from explicit to implicit, from hand-designed to automatical-learned, which can be followed to design your own innovative approaches!

 

Occlusion

1. Hirerarchical part model for visibility estimation [1][5]
2. Occlusion data or feature augmentation [2] see Data and Feature Augmentation

Scale

［14］

1. Image pyramid: compute feature from each level of the image pyramid [1]
1. computational expensive, usually applied during the inference stage
2. Encoder-decoder: feature map from multiple convolutional and deconvolutional layers 
1. pyramidal feature hierarchy [5] [6] [7] 
2. feature pyramid [11]
3. Deeper with atrous convolution
4. Spatial pyramid pooling[14]
1. 2,3,4 can be combined, and are explored in [15]
5. Attemtion
1. Detect and focus on a smaller region in each stage [2]

Data Imbalance 

1. Emphasize on balanced compilations of datasets in the first place
1. Collecting their samples approximately uniformly
2. Data and Feature Augmentation
1. Dropout 1/2 neurals for better generalization [4]
2. GAN
1. Generating hard feature maps for occlusion and deformation in object detection task [2]
2. Conducting over-sampling of minority classes or under-sampling from the majority classes
1. Weakness
1. change the underlying data distributions and may result in suboptimal exploitation of available data
2. increased computational effort and/or risk of over-fitting when repeatedly visiting the same samples
1. SMOTE and derived variants on ways to avoid over-fitting
3. Data Mining for Hard Examples
1. Online hard example mining (OHEM) [3] for both intra-class data imbalance and positive-negative imbalance 
4. Cost-sensitive learning
1. Focal loss [8]: greater loss for harder example
2. Loss Max-Pooling for Semantic Image Segmentation [16]：by the maximization with respect to pixel weighting functions, the loss function providing an adaptive re-weighting of the contributions of each pixel. Pixels incurring higher losses during training are weighted more than pixels with a lower loss. 

Local & global information combination

1. Deep learning can learn some multi-scale information automatically[9]
2. Top-down semantic from FPN (focus on each scale) [11]
3. Multi-scale combination & selection from GBD-Net (focus on combination of scales) [12]

Utilization of Context

1. In traidition machine learning, mainly used as  refine object scores [1] [10]
2. RNN [13]

Utilization of Object Part Information

1. DPM [1]
2. Deep learning with DPM [5]
3. Position sensitive ROI pooling [10]: construct a score map from different channels (results of part detectors) of feature map

Metric learning

Mainly used in recognition.

Computation Efficiency

1. eliminate redudant layers 
2. spatial adaptive computation

Reference:

1. Felzenszwalb, Pedro F., et al. "Object detection with discriminatively trained part-based models." IEEE transactions on pattern analysis and machine intelligence 32.9 (2010): 1627-1645.
2. Wang, Xiaolong, Abhinav Shrivastava, and Abhinav Gupta. "A-fast-rcnn: Hard positive generation via adversary for object detection." arXiv preprint arXiv:1704.03414 (2017).
3. Shrivastava, Abhinav, Abhinav Gupta, and Ross Girshick. "Training region-based object detectors with online hard example mining." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016.
4. Srivastava, Nitish, et al. "Dropout: a simple way to prevent neural networks from overfitting." Journal of Machine Learning Research 15.1 (2014): 1929-1958.
5. Ouyang, Wanli, and Xiaogang Wang. "Joint deep learning for pedestrian detection." Proceedings of the IEEE International Conference on Computer Vision. 2013.
6. Liu, Wei, et al. "Ssd: Single shot multibox detector." European conference on computer vision. Springer, Cham, 2016.
7. Fu, Cheng-Yang, et al. "DSSD: Deconvolutional Single Shot Detector." arXiv preprint arXiv:1701.06659 (2017).
8. Tsung-Yi Lin, etal. "Focal Loss for Dense Object Detection.” IEEE International Conference on Computer Vision (ICCV), 2017
9. Girshick, Ross. "Fast r-cnn." Proceedings of the IEEE international conference on computer vision. 2015.
10. C. Galleguillos and S. Belongie. Context based object cate- gorization: A critical survey. In CVPR, 2010. 1, 2
11. Lin T Y, Dollár P, Girshick R, et al. Feature pyramid networks for object detection[J]. arXiv preprint arXiv:1612.03144, 2016.
12. Zeng X, Ouyang W, Yang B, et al. Gated Bi-directional CNN for Object Detection[C]// European Conference on Computer Vision. Springer, Cham, 2016:354-369.
13. Bell S, Lawrence Zitnick C, Bala K, et al. Inside-outside net: Detecting objects in context with skip pooling and recurrent neural networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016: 2874-2883.
14. Rethinking Atrous Convolution for Semantic Image Segmentation (DeepLab v3)
15. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation (DeepLab v3+)
16. Loss Max-Pooling for Semantic Image Segmentation [16]