Skip to content

Latest commit

 

History

History
34 lines (34 loc) · 2.04 KB

R-CNN.md

File metadata and controls

34 lines (34 loc) · 2.04 KB
  • #paper/to-read ~ [[2013 CE]] ~ [[Object Detection]], [[Semantic Segmentation]]
    • Rich feature hierarchies for accurate object detection and semantic segmentation
    • https://arxiv.org/abs/1311.2524
    • Sequel papers:
      • [[Fast R-CNN]]
      • [[Faster R-CNN]]
      • [[Mask R-CNN]]
      • [[Context R-CNN]]
    • Mentioned papers:
      • [[Regionlets]]
      • [[Objectness of Image Windows]]
      • [[Recognition Using Regions]]
      • [[Category Independent Object Proposals]]
      • [[DeCAF]]
      • [[Adaptive Deconvolutional Networks]]
      • [[Diagnosing Error in Object Detectors]]
    • Mentioned topics:
      • [[Deformable Part Models]]
      • [[Hard Negative Mining]]

  • Summary

    • Architecture: input image -> extracting region proposals (~2k) -> feature-extraction [[Convolutional Neural Network, CNN|CNN]] -> class prediction or bounding box regression [[Support Vector Machine, SVM|SVM]] (or both).
      • [[Intersection Over Union, IoU]] with ground truth labels should be greater than 0.5.
      • At test time, [[Intersection Over Union, IoU|IoU]] is used to select the dominating prediction for a region.
      • With the enabled bounding box [[Regression]], the model learns to predict shifts of the predicted bounding box relative to the coordinates of the proposed region.
    • R-CNN can scale to thousands of classes without resorting to hierarchical tricks or [[Hash Function|Hashing]].
    • [[Training Process|Training]] process consists of two parts.
      • Supervised pre-training happens on a big [[Classification]] dataset to make the CNN part learn without overfitting on (usually) small detection datasets.
      • Domain-specific fine-tuning is used to train SVM heads on region proposals.
        • Those proposals should fit $227 \times 227$ pixels so they are usually warped.
    • Different CNN architectures yield different [[Mean Average Precision, mAP|mAPs]] but also require different amount of [[Compute]].
    • For region proposals, [[Selective Search]] is used but other methods are appropriate too.
    • For [[Semantic Segmentation]], [[Constrained Parametric Min-Cuts, CPMC]] is used.

![[r-cnn.pdf]]