Weakly Supervised Deep Detection Networks 阅读笔记

最新推荐文章于 2022-11-14 00:13:39 发布

shannanyinxiang

最新推荐文章于 2022-11-14 00:13:39 发布

阅读量8.3k

点赞数

CC 4.0 BY-SA版权

分类专栏：论文阅读文章标签： deep-learning weakly supervised

本文链接：https://blog.youkuaiyun.com/qq_32858815/article/details/77951859

论文阅读专栏收录该内容

2 篇文章

订阅专栏

本文介绍了Weakly Supervised Deep Detection Networks的整体架构，包括预训练网络、区域级描述符、识别与检测分支，以及训练过程中的空间正则化。与其它方法对比，WSDDN的检测分支独立于识别分支，更侧重于设计。实验部分展示了WSDDN在CorLoc上的表现，尽管存在将多个对象实例归为一个框的问题，但仍然展示了其在弱监督检测中的潜力。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

Weakly Supervised Deep Detection Networks 阅读笔记

Overall architecture

1. Existing network(such as AlexNet pre-trained on ImageNet）

2. SPP --> region level descriptor

3. (1) class score --> recognition

(2) probability distribution(which region contains the most salient image structure) --> detection

4. aggregate the recognition and detection scores to predict the class of image(image level supervision)

Compared with other method

1. MIL: Use the appearance model itself to perform region selection

WSDDN: detection branch is independent of recognition branch

2. Bilinear architecture: two streams are symmetric

WSDDN: detection branch is explicitly designed

Method

1. Pre-trained network

2. Weakly supervised deep detection network

(1) Region level descriptor:

Region proposal: SSW, EB

(2) Classification data stream: fc + softmax

(3) Detection data stream: fc + softmax(differently defined)

(4) Combined region scores and detection

Final score of each region:

Then rank regions for each class independently.

Then apply nms(0.4)

(5) Image-level classification scores

Image level class score:

(yc in (0, 1))

Softmax is not applied because one image can have multiple label

3. Training WSDDN

A collection of images x_i, i=1, 2, … , n

Image level labels y_i∈ {-1, 1}^C

4. Spatial regulariser

Penalize the feature map discrepancies between the highest scoring region and the regions with at least 60% IoU during training.

Experiments

CorLoc: the percentage of images that contain at least one instance of the target object class for which the most confident detected bounding box overlaps by at least 0.5 with one of these instances.

Problem: (1) group multiple object instances with a single bounding box

(2)focus on parts rather than the whole object

Result: