论文笔记之网络结构篇:RCNN

PAST: Combine multiplle low-level image features with high-level context

Key insights:

1. CNN ---- bottom-up region proposals in order to localize and segment objects
2. labeled training data is scare ---- supervised pre-training for anauxiliary task, followed by domain-specific fine-tuning, yields a significant performance boost.

The conventional solution to this problem is to use unsupervised pre-training,followed by supervised fine-tuning .
supervised pre-training on a large auxiliary dataset(ILSVRC),followed by domain specific fine-tuning on a small dataset (PASCAL), is an effective paradigm for learning high-capacity CNNs when data is scarce.

Object detection with R-CNN

Three module:

1. generates category-independent region proposals
2. a large convolutional neural network that extracts a fixed-length feature vector from each region
3. a set of class specific linear SVMs.

2.1 Module design

Region proposals:

selective search

Feature extraction

2.2 Test-time detection

proposal 2000 region proposals ——> for each class, score each extracted feature vector --> non-maximum suppression

Run-time analysis

efficient

• all CNN parameters are shared across all categories

• the feature vectors computed by the CNN are low-dimensional

2.3 Training

Supervised pre-training

pre-trained the CNN on a large auxiliary dataset (ILSVRC2012 classification) using image-level annotations only (boundingbox labels are not available for this data).

Domain-specific fine-tuning

replacing the CNN’s ImageNetspecific 1000-way classification layer with a randomly initialized (N + 1)-way classification layer (where N is the number of object classes, plus 1 for background)

Object category classifiers – overlap

2.4 Results on PASCAL VOC 2010-12

2.5 Results on ILSVRC2013 detection (one with and one without bounding-box regression)