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
| Network | Top-1 | Top-5 | sha256sum | Architecture |
|---|---|---|---|---|
| MobileNet v1 | 70.81 | 89.85 | 8d6edcd3 (16.2 MB) | netscope |
| MobileNet v2 | 71.90 | 90.49 | a3124ce7 (13.5 MB) | netscope |
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:
- https://arxiv.org/abs/1506.01497
- https://www.arxiv.org/pdf/1608.08021v3.pdf
- https://arxiv.org/abs/1605.06409
train data 110 thousand iterations Mean AP = 0.7498