object detection and recognition

Post on 22-Jan-2018

1.498 Views

Category:

Technology

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Proprietary and confidential. Do not distribute.

using neon for object detection and localization

Yinyin Liu, PhD March 3 2016

Outline

2

Intro to Deep Learning• From a user’s perspective, how to use neon to solve your

problem

• Use object localization problem as an example to

• understand and utilize neon architecture

• implement a new model

Object detection and localization

3

Object detection and localization

4

Fast region-based CNN (RCNN)

5

Fast R-CNN [Girshick 2015] http://arxiv.org/abs/1504.08083

• Pre-trained the ConvNets

• ROI pooling

• Branch architecture

• cost function to consider both

functions

neon architecture

6

dataset

model

Optimizer

Cost

callbacks

Layers/container

Metric

Architecture

Building fast R-CNN network

7

model

Optimizer

smooth L1 cost

callbacks

loading trained VGG layers

Branch architecture

Pooling

ROI pooling

Dropout

Affine

Conv

data iterator

PASCAL VOC

Object detection metric

• Input • Image • ROIs

• Target • class label • box regression • box regression mask

new components that were not in neon

PASCAL VOC in a dataset container

8

PASCAL VOC

• Input • Image • ROIs

• Target • class label • box regression • box regression mask

ROI pooling layer

9

• ROI pooling layer combines feature map from a

layer, and ROIs from the dataset

• Make a ROI pooling a container

• contains the ConvNet layers

• interface with dataset directly

ROI pooling layer

10

• Any new layer or container needs to:

• work as part of the model’s forward and

backward propagation process

• have fprop and bprop functions

• start from an python implementation

• write backend (GPU) support for speed

Cost and metric

11

• new type of cost need to be derived from Cost

class

• new type of metric need to be derived from

Metric class

Fast R-CNN model

12

Fast R-CNN model

13

Trained and tested in neon

14

Trained and tested in neon

15

Trained and tested in neon

16

Contact

17

yinyin@nervanasys.com

github.com/NervanaSystems/neon

\

top related