[ICCV2017] SubUNets from Oscar Koller team

论文链接：http://openaccess.thecvf.com/content_ICCV_2017/papers/Camgoz_SubUNets_End-To-End_Hand_ICCV_2017_paper.pdf

源代码：https://github.com/neccam/SubUNets

一、网络结构概览

二、网络各模块简介

1. CNN

CaffeNet, use the weights that were pre-trained on ImageNet.

2. Hand SubUNet

We use the One-Million Hands dataset for training the Hand SubUNet.

It has over 1.2 million hand images, from which 1 million images were labelled with one of 60 hand shape classes.

3. Word SubUNet

 

 

三、各模块联合

 两种训练策略

1. Fixed

We pre-train two SubUNets depicted in Figure4, blue blocks are pre-trained and fixed, green blocks are removed, while white block are trained for the task.

2. Not Fixed

All weights are trained using the gradients produced by all three loss layers.

 最终选用 Not Fixed 训练策略

四、解码策略

 1. Beam search - Full Sum VS. Greedy - Viterbi

 2. Different topologies

 [26] is CVIU2015 from Koller, [27] is CVPR2016 from Koller