Real - Time Face Recognition on Jetson Tx2 using TensorRT Tamas - - PowerPoint PPT Presentation
Real - Time Face Recognition on Jetson Tx2 using TensorRT Tamas Grobler 11 . 10 . 2017 GTC Table of Contents The The Problem Results Conclusion solution Real Time Video Recognition Testing Method Real Time Analysis Framework Recognition
Tamas Grobler 11.10.2017 GTC
Real-Time Face Recognition
Real Time Video Analysis Face Detection Face Recognition
The Problem
Recognition Framework Training Method
The solution
Testing Method Recognition results Inference time comparison Real Time Considerations Video
Conclusion
01
Results
Real-Time Face Recognition
02
Real-Time Face Recognition
03
Real-Time Face Recognition
04
Region-based CNN methodology
Backbone: PVANet with HyperNet-inspired hyper- features Region Proposal Network from Faster R-CNN Network Scales with Image Size
Training:
Non Maximum Suppression Backbone pre-trained on ImageNet Hard Negative Mining 05
Real-Time Face Recognition
Ultinous film head corpus + subset of HollywoodHeads dataset
Base Model: GoogLeNet Multiple crops of each image (160x160 -> 144x144) Frontalization using Spatial Transformer Network (144x144 -> 128x128) Training
MsCeleb dataset: ~10M images from ~100k classes Augmentation: random crops, mirroring Trained for both Classification and Discrimination (Triplet Loss) with Hard Negative Mining
Real-Time Face Recognition
06
Test Data: NIST IJB-A Dataset
Average over 10 splits 1000 positive/10000 negative pairs per split 07
Real-Time Face Recognition
NVIDIA Pascal architecture 3840 NVIDIA CUDA Cores 12 GB memory; 547.7 GB/s NVIDIA Pascal architecture 256 NVIDIA CUDA Cores 8 GB memory; 58.3 GB/s
NVIDIA Titan Xp
Caffe NVIDIA TensorRT
NVIDIA Jetson TX2
08
Real-Time Face Recognition
NVIDIA CUDA 8 NVIDIA cuDNN 7 3.0 Release Candidate 32b, 16b, 8b arithmetic
09
Real-Time Face Recognition
10
Real-Time Face Recognition
Frame rate: 10 fps -> 100 ms for detection + recognition Head Detection time per 1536x864 frame (speculative for 16 bit): < 30 ms Face recognition can handle 50 images in ~66 ms 5 crops per face This allows for 10 simultaneous recognitions per frame (30 ms + 66 ms) Future work: test TensorRT with Int8 arithmetic for both accuracy and inference time 11
Real-Time Face Recognition
Real-Time Considerations for Jetson TX2 with TensorRT (16-bit arithmetic)
Real-Time Face Recognition
12
Real-Time Face Recognition
13 György Balogh János Locki József Németh Attila Szabó tamas.grobler@ultinous.com ultinous.com
[1] arXiv:1611.08588 PVANet: Lightweight Deep Neural Networks for Real-time Object Detection Sanghoon Hong, Byungseok Roh, Kye-Hyeon Kim, Yeongjae Cheon, Minje Park [2] arXiv:1604.00600 HyperNet: Towards Accurate Region Proposal Generation and Joint Object Detection Tao Kong, Anbang Yao, Yurong Chen, Fuchun Sun [3] arXiv:1506.01497 Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun [4] arXiv:1409.4842 Going deeper with convolutions Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erha, Vincent Vanhoucke, Andrew Rabinovich [5] arXiv:1701.07174 Towards End-to-End Face Recognition through Alignment Learning Yuanyi Zhong, Jiansheng Chen, Bo Huang [6] “This product contains or makes use of the following data made available by the Intelligence Advanced Research Projects Activity (IARPA): IARPA Janus Benchmark A (IJB-A) data detailed at Face Challenges homepage." [7] arXiv:1408.5093 Caffe: Convolutional Architecture for Fast Feature Embedding Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, Trevor Darrell
Real-Time Face Recognition
14