Skip to content

Latest commit

 

History

History
112 lines (60 loc) · 3.01 KB

README.md

File metadata and controls

112 lines (60 loc) · 3.01 KB

RCNN (Ongoing)

Tensorflow implementation of the RCNN object detection system as proposed by Rich feature hierarchies for accurate object detection and semantic segmentation .

Test Notebook: Binder

Documented Notebook: Binder

Pascal VOC2012

Number of Objects per Image

Frequency Distribution of Classes

Sample Images + Ground Truth

RCNN

Overview

The RCNN system was proposed by Ross Girshick, Jeff Donahue, Trevor Darrell and **Jitendra Malik from UC Berkeley in their paper Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation. RCNN stands for Regions with CNN features, which summarizes the working of the system in very simple terms, generating region proposals with classification using CNNs. The RCNN consists of 3 simple stages:

  1. Given an input image, around 2000 bottom-up region proposals are extracted.

  2. Computation of features for each proposal using a large convolutional neural network (like pre-trained VGG or ResNets).

  3. Classification of each region using class-specific linear SVMs (or MLPs).

**

Seletive Search

For generating the region proposals, we would look towards the following 2 papers:

  1. Efficient Graph-Based Image Segmentation proposed by by Pedro F. Felzenszwalb and Daniel P. Huttenlocher.

  2. Selective Search for Object Recognition proposed by J R R Uijlings, K E A van de Sande, T Gevers and A W M Smeulders.

Felzenszwalb Segmentation

Texture Gradients

Generating Bounding Box Proposals

Narrowing down on objects using IoU

Narrowing down on objects using IoU (OpenCV Implementation)

Citation

@misc{1311.2524,
    Author = {Ross Girshick and Jeff Donahue and Trevor Darrell and Jitendra Malik},
    Title = {Rich feature hierarchies for accurate object detection and semantic segmentation},
    Year = {2013},
    Eprint = {arXiv:1311.2524},
}