DeepBox: Learning Objectness with Convolutional Networks

ICCV 2015
code https://github.com/weichengkuo/DeepBox

proposal re-ranker ,本文是对候选区域重新排序。从 Edge box的2000个结果中选出500个好的候选区域,怎么选了? 使用一个小的 CNN网络

DeepBox: Learning Objectness with Convolutional Networks_第1张图片

3.1. Network Architecture
our network architecture is:
conv(11,96,4)−pool(3,2)−conv(5,256,1)−fc(1024)−fc(2).

3.2. Sharing computation for faster reranking
采用SPP方法,对图像一次性计算卷积特征得到 shared convolutional feature maps。这样每个候选区域的特征就不用单独计算了,直接获取

3.3. Training Procedure
3.3.1 Initialization
前两个卷积层使用 Imagenet model 参数初始化
3.3.2 Training on Sliding Windows 先容易样本
3.3.3 Training on Hard-negatives 后困难样本

4 Experiments
因为该模型可以解决小物体,而COCO数据库上小物体比较多,所以在 COCO数据库上测试

DeepBox: Learning Objectness with Convolutional Networks_第2张图片

感觉有了 Faster R-CNN, 本文有点多余啊!

你可能感兴趣的:(DeepBox: Learning Objectness with Convolutional Networks)