for Object Detection Forrest Iandola Ning Zhang Ross Girshick Trevor Darrell and Kurt Keutzer Deformable Parts Model DPM state of the art algorithm for object detection 1 Several attempts to accelerate multicategory DPM detection such as 2 3 ID: 473242
Download Presentation The PPT/PDF document "Scaling Up Deformable Parts Models (DPMs..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Scaling Up Deformable Parts Models (DPMs) for Object Detection
Forrest Iandola, Ning Zhang, Ross Girshick, Trevor Darrell, and Kurt Keutzer
Deformable Parts Model (DPM): state of the art algorithm for object detection [1]Several attempts to accelerate multi-category DPM detection, such as [2] [3]Our goal: fast single-category detection in videosWant more categories? Combine our work with [2] or [3]. Exploring existing and new implementations
Deformable Parts Model Detection
P. Felzenszwalb, R. Girshick, D. McAllester, D. Ramanan. Object Detection with Discriminatively Trained Part Based Models. PAMI, 2010.H. Song, R. Girshick, T. Darrell. Discriminatively Activated Sparselets. ICML, 2013.T. Dean, et al. (Google, Inc.) Fast, Accurate Detection of 100,000 Object Classes on a Single Machine. CVPR, 2013.C. Dubout and F. Fleuret. Exact Acceleration of Linear Object Detectors. ECCV, 2012.F. Iandola, D. Sheffield, M. Anderson, P. Phothilimthana, and K. Keutzer. Communication-Minimizing 2D Convolution in GPU Registers. ICIP, 2013.H. Niknejad et al. On-Road Multivehicle Tracking Using Deformable Object Model and Particle Filter With Improved Likelihood Estimation. IEEE Trans. Intell. Transportation. 2012.M. Andriluka et al. Vision based victim detection from unmanned aerial vehicles. IROS, 2010.
Image convolution
Image credit: Apple
Introduction
Convolution Tutorial
Edge detection example
Fast Convolution for DPMs
Image
Horizontal Edges
References
Filter
{
forresti
,
nzhang
,
rbg
,
trevor
, keutzer}@eecs.berkeley.edu
Input Image
Extract HOG feature descriptors
Feature matching with learned model
Distance transform and scoring
Structured gridHistogram andConvolution
Structured gridConvolution with several filters
Output Detections
Vector distance,
Structured grid
Na
ive vs.
F
ast
GPU code
Prefetch
image feature windows to registers (more details: [5])
Prefetch
filters to shared memory
Reuse image feature data when convolving with several filters
Convolution global memory reads
Naive approach without data reuse: (HOG image size) * (filter size) * (#of filters)Lower bound:(HOG image size) + (filter size) * (#of filters)
Feature Matching
(Convolution), 62%
Distance
Transform16%
HOG FeatureExtraction22%
Performance breakdown of FFT-based DPM detection [4]
5 fps at 640x480
GPU:
NVIDIA GTX680CPU: Intel i7-3930k
DPM Applications
Traffic safety [6]
UAV pedestrian detection and activity recognition [7]
Vehicle Classification
Image credit: Toyota Technical Institute
Image credit:
TU Darmstadt
Predicted: 2000 AM General Hummer
Actual:
2000 AM General
Hummer
Predicted: 2012 BMW 3-series
Actual: 2012
BMW 3-
series
Predicted: 2012 GMC
Savana
Actual: 2012
Chevrolet
Express
Badge-engineered twins!