Fast Detection of Multiple Textureless 3-D Objects
Hongping Cai, Tomas Werner and Jiri Matas
Center for Machine Perception, Czech Technical University, Prague
Introduction
Results
Stage1: Fast Hypotheses Pruning
LAB−8objs, Table−3pt
A novel sparse edge-based image descriptor
Task: Detection+Recognition +Pose Estimation of Textureless objects
...
...
...
...
...
...
Test image
φI (p)
p
m reference points (in red) are assigned two features:
(dI (p1), . . . , dI (pm), φI (p1), . . . , φI (pm)).
I dI (p): distance to the nearest edge.
I φI (p): orientation of the nearest edge.
screw:0.84
bridge:0.92
lid:0.86
...
...
I
dI (p)
screw:0.77
Training templates
I
Detection Rate
driver:0.80
0
0
I
0.5
1
1.5
2
2.5
False Positive Per Image
3
Quantize the descriptors
I Put all training template indices
into a table
I Fast voting scheme
I Speed up by grouping triplets of
reference points
Texture-free objects:
Appearing in both industrial manipulation and daily life.
I Approaches with local detectors and descriptors fail.
We use edges.
I
3-D pose:
Large shape variety from different viewpoints.
Obj30 [Damen et al. (2012)]
I 30 objects
I 7,056 training templates
I 1,200 test images of size 640×480
I Prec=0.85@recall=0.5
I 0.26 s/frame
0.8
I
I
OrigEdge, CompOCM (DR:0.66)
SelEdge, CompOCM (DR:0.74)
OrigEdge, OrigOCM (DR:0.51)
SelEdge, OrigOCM (DR:0.63)
1
Challenges
I
0.4
0.2
Fast voting with quantized features
I
0.6
Obj30, Table−3pt
We propose a fast edge-based approach for detection and approximate
pose estimation of multiple textureless objects in a single image, by
combining an efficient voting procedure with an effective verification stage.
I
CMP-8objs*
I 8 objects
I 12,960 training templates
I 60 test images of size 640×480
I DR=0.74@FPPI=1.0
I 0.65 s/frame
0.8
precision
I
1
0.6
0.4
OrigEdge, CompOCM (pre:0.74)
SelEdge, CompOCM (pre:0.85)
OrigEdge, OrigOCM (pre:0.54)
SelEdge, OrigOCM (pre:0.54)
Damen et al. (2012) (pre:74%)
0.2
0
0
0.2
0.4
0.6
0.8
1
recall
lid
eye
whiteblock
cup
whiteblock
Real-time requirement:
I
A key challenge due to a large set of templates.
Stage2: Improved Oriented Chamfer Matching (OCM)
block
lid
lid
driver
Acquiring Training Templates
block
eye
screw
driver
driver
bridge whiteblock
cup
lid
Acquiring training set for our CMP-8objs dataset. Left: each object is placed
on a turntable and captured by video camera, covering a viewing hemisphere.
Right: examples of 12,960 training edge-based templates.
Compensating the bias towards simpler shapes
I Our CompOCM score:
e ∈ T dI (e) ≤ θd , |φT (e) − φI (e)|π ≤ θφ ,
(1)
sλ(T , I) =
λ|T | + (1 − λ)|T |
P
n
1
|T | = n i=1 |Ti |: the average number of edges over all training templates, λ ∈ [0, 1].
I Standard OCM score: s1(T , I)
Two-stage cascade
I
Driver:0.54
Hammer:0.60
E−driver:0.60
Headphone:0.83
Box:0.77
Pliers:0.71
BookLight:0.62
TapeDispenser:0.64
Mug:0.95
Wood:0.58
Mouse:0.57
StarModel:0.56
Wrench:0.55
Hammer:0.51
Mug:0.73
Tape:0.53
Apple
Selection by edge orientations.
I
screw
Only the edges with higher repeatability after slight changes of viewpoint are kept.
40% of edges corresponding to the two highest orientation bins are randomly removed.
#edge:392
#edge:277
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
pi/12
pi/4
5pi/12 7pi/12
3pi/4 11pi/12
Left: original edges and its orientation histogram,
Right: selected edges and its orientation histogram.
* http://cmp.felk.cvut.cz/data/textureless
eye
Selecting discriminative edges of templates
I Selection by stability to viewpoint.
I
Our Approach
screw
Email: [email protected]
Example detections for CMP-8objs (top) and Obj30 (bottom) datasets. TP, FP
and FN are shown in green, red and yellow, respectively.
Summary
Fast detection of textureless 3D objects, better than state-of-the-art accuracy.
I Speed: 4fps on Obj30 and 1.5fps on CMP-8objs dataset.
I Purely edge-based method, no other clues, e.g., color, used.
I A new dataset CMP-8objs is generated for evaluation of 3-D textureless
object detection and pose estimation.*
I
pi/12
pi/4
5pi/12 7pi/12 3pi/4 11pi/12
Homepage: http://cmp.felk.cvut.cz/∼caihongp/