1. Progress Report
2019/1/3
PHHung

2. Previous Deep learning … The promised land for CV ?
2019/1/3
Deep learning … The promised land for CV ?
Implement convolutional neural network step by step (follow by UFLDL tutorial)
Back propagation Algorithm
Sparse Autoencoder
Softmax Regression
Stacked Autoencoders
Convolutional neural network

3. Two direction State of the art algorithm survey
2019/1/3
State of the art algorithm survey
2 hidden layers for pedestrian detection , CVPR 2013
22 hidden layers for class object detection , CVPR 2015?
Hardware implementation survey
Works from 332
TeraDeep from Prudue
FPGA & ASIC implement from NYU (LeCun)

4. A small toy
2019/1/3
Use the building boxes from previous to build a pedestrian detector
Train on INRIA pedestrian dataset
Test on Caltech pedestrian dataset
Classifier + Sliding window
Very slow… I should try objectness filter instead of sliding window next time
2 hidden layers (200/100nodes )
Result is approximately at here
FPPI~1.2
Miss rate~12.4%

5. Back to reality… Why we need analysis engine?
2019/1/3
Why we need analysis engine?
=> To save power consumption
For ReSSP : 197mW in VGA , 39.7mW in QCIF
For Deep learning engine :
HMAX
ASIC: 205mW in 256x256 (Neocortical from 332 , ISSCC 12)
ConNet
FPGA: 10W in 500x375 (NeuFlow from NYU LeCun , CVPR 11)
ASIC: 580mW in 500x375 (NeuFlow from Purdue & NYU , MWCAS 12)
FPGA: 4W (TeraDeep from Purdue , CVPR 14)
None of above can have power lower than 100mW…
How about power & detection rate at low resolution (QCIF) ?