Poster - Forrest Iandola

Download Report

Transcript Poster - Forrest Iandola

Scaling Up Deformable Parts Models (DPMs)
for Object Detection
Forrest Iandola, Ning Zhang, Ross Girshick, Trevor Darrell, and Kurt Keutzer
{forresti, nzhang, rbg, trevor, keutzer}@eecs.berkeley.edu
Introduction




Deformable Parts Model Detection
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
videos
 Want more categories? Combine our work
with [2] or [3].
Exploring existing and new implementations
Input Image
Extract HOG
feature descriptors
Structured grid
Histogram and
Convolution
Convolution Tutorial
Image convolution
Distance
transform and
scoring
Feature matching
with learned
model
Structured grid
Convolution with
several filters
Output Detections
Vector distance,
Structured grid
Fast Convolution for DPMs
DPM Applications
Traffic safety [6]
Naive vs. Fast 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
Distance HOG
Transform Feature
16% Extraction
22%
Feature Matching
(Convolution), 62%
Convolution global memory reads
Naive approach without data reuse:
(HOG image size) * (filter size) * (#of filters)
Image credit: Apple
Image credit: Toyota Technical Institute
Performance breakdown of
FFT-based DPM detection [4]
Edge detection example
5 fps at 640x480
Vehicle Classification
500
DPM Convolution
CPU Peak
Communication-avoiding GPU direct convolution (ours)
Naive GPU direct convolution (ours)
CPU FFTW [6] (relative to direct convolution)
CPU DPM voc-release5 direct convolution [11]
00
0 x6 0
800x48
64
00
x12
20
00
19
x12
00
16
60
25
x16
00
Image credit: TU Darmstadt
400
1. P. Felzenszwalb, R. Girshick, D. McAllester, D. Ramanan. Object Detection with
Discriminatively Trained Part Based Models. PAMI, 2010.
2. H. Song, R. Girshick, T. Darrell. Discriminatively Activated Sparselets. ICML, 2013.
3. T. Dean, et al. (Google, Inc.) Fast, Accurate Detection of 100,000 Object Classes on a
Single Machine. CVPR, 2013.
4. C. Dubout and F. Fleuret. Exact Acceleration of Linear Object Detectors. ECCV, 2012.
5. F. Iandola, D. Sheffield, M. Anderson, P. Phothilimthana, and K. Keutzer.
Communication-Minimizing 2D Convolution in GPU Registers. ICIP, 2013.
6. H. Niknejad et al. On-Road Multivehicle Tracking Using Deformable Object Model and
Particle Filter With Improved Likelihood Estimation. IEEE Trans. Intell. Transportation.
2012.
7. M. Andriluka et al. Vision based victim detection from unmanned aerial vehicles. IROS,
2010.
UAV pedestrian detection and activity recognition [7]
GFLOPS/s
300
References
GPU: NVIDIA GTX680
CPU: Intel i7-3930k
200
Horizontal Edges
100
0
Filter
Predicted: 2012 BMW 3-series Predicted: 2012 GMC Savana
Actual: 2012 BMW 3-series
Actual: 2012 Chevrolet Express
Badge-engineered twins!
Image size
Image
Predicted: 2000 AM General Hummer
Actual: 2000 AM General Hummer
Lower bound:
(HOG image size) + (filter size) * (#of filters)