Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
Fast Multi-frame Stereo Scene Flow with Motion Segmentation (CVPR 2017)
1. 1
Fast Multi-frame Stereo Scene Flow
with Motion Segmentation
Tatsunori Taniai*
RIKEN AIP
Sudipta N. Sinha
Microsoft Research
Yoichi Sato
The University of Tokyo
CVPR 2017 Paper
* Work done during internship at Microsoft Research and partly at the University of Tokyo.
3. 3
Contributions
• New unified framework
– Stereo (depth / disparity)
– Optical flow (2D motion field)
– Motion segmentation (binary mask of moving objects)
– Visual odometry (6 DoF camera ego-motion)
In our framework
• Result of each task benefits others, leading to higher accuracy and efficiency
• Joint task is decomposed into simple optimization problems (in contrast to
existing joint methods)
Results
• Accurate: achieved 3rd rank on KITTI benchmark
• Fast: 10~1000x faster than state-of-the-art methods
4. 5
Scene Flow: Problem Definition
𝑿 𝑡 = 𝑥 𝑡, 𝑦𝑡, 𝑧𝑡
𝑝𝑝′
𝐼𝑡
0
𝐼𝑡
1
Stereo disparity
1D horizontal translation
by object depth 𝑧
𝑝′
5. 6
Scene Flow: Problem Definition
𝐼𝑡
0
𝐼𝑡+1
0
𝐼𝑡
1
𝐼𝑡+1
1
𝑿 𝑡
𝑝
𝑝′
Optical flow
2D translation
by camera and
object motions
𝑝′
𝑿 𝑡+1
6. 7
Scene Flow: Problem Definition
𝑿 𝑡
Stereo disparity
1D horizontal translation
by object depth 𝑧
𝐼𝑡
0
𝐼𝑡+1
0
𝑿 𝑡+1
Optical flow
2D translation
by camera and
object motions
All together implicitly
represent 3D
motions of points
7. 8
Applications
Autonomous driving
[Menze+ CVPR 15]
Action recognition
[Wang+ CVPR 11]
Depth and flow map sequences are useful in many applications
But optical flow estimation is VERY SLOW.
9. 10
Optical Flow vs Stereo
Optical flow Stereo matching
1D translation 2D translationSearch space
Motion factor Object motion,
Ego-motion, etc.
Object depth
Optical flow is much more difficult & expensive than stereo
10. 11
Dominant Rigid Scene Assumption
Most of the points are static.
Their flows are due to camera motions.
11. 12
Flow Estimation by Depth and Camera Motion
𝐼𝑡
Rigid flow map Ground truth flow map Error map of rigid flow
Surface 𝐷𝑡
𝐼𝑡+1
Surface 𝐷𝑡+1
𝐼𝑡 𝐼𝑡+1
Given rigid flow map,
we only need to
recompute flow for
moving objects.