pva-mobilenet-v2

Introduction

This is a Caffe implementation of Google's MobileNets (v1 and v2). For details, please read the following papers:

• [v1] MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

• [v2] Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation

try using pvanet architecture to trained voc data

Pretrained Models on ImageNet using https://github.com/shicai/MobileNet-Caffe.git

accuary

AP for aeroplane = 0.5731 AP for bicycle = 0.6798 AP for bird = 0.4748 AP for boat = 0.4037 AP for bottle = 0.3515 AP for bus = 0.7517 AP for car = 0.7366 AP for cat = 0.7646 AP for chair = 0.2972 AP for cow = 0.3399 AP for diningtable = 0.5541 AP for dog = 0.7086 AP for horse = 0.7416 AP for motorbike = 0.7236 AP for person = 0.7001 AP for pottedplant = 0.2407 AP for sheep = 0.0702 AP for sofa = 0.5284 AP for train = 0.6882 AP for tvmonitor = 0.4034 Mean AP = 0.5366

#pva-mobilenet-v2 model: url: https://pan.baidu.com/s/1Cl4MXiU7otkB6bxGIOLxwg

ohem Introduction:

This implemention is using rfcn ohem way train voc dataset, base work is pvanet, following papers:

1. https://arxiv.org/abs/1506.01497

2. https://www.arxiv.org/pdf/1608.08021v3.pdf

3. https://arxiv.org/abs/1605.06409

train data 110 thousand iterations Mean AP = 0.7498