| <Poster Width="1942" Height="883"> |
| <Panel left="18" right="129" width="465" height="411"> |
| <Text>Introduction</Text> |
| <Text>• Some object classes are hard to reconstruct</Text> |
| <Text>– Lack of texture</Text> |
| <Text>– Transparency</Text> |
| <Text>– Reflection</Text> |
| <Text>• Solution: shape prior</Text> |
| <Text>– Shapes within object class similar</Text> |
| <Text>– Local distribution of surface normals</Text> |
| <Figure left="59" right="349" width="116" height="87" no="1" OriWidth="0.102312" OriHeight="0.0598749 |
| " /> |
| <Figure left="56" right="450" width="127" height="90" no="2" OriWidth="0.107514" OriHeight="0.0580876 |
| " /> |
| <Figure left="184" right="344" width="250" height="193" no="3" OriWidth="0.262428" OriHeight="0.152815 |
| " /> |
| </Panel> |
|
|
| <Panel left="18" right="550" width="464" height="305"> |
| <Text>Formulation</Text> |
| <Text>• Baseline Method: Volumetric depth map fusion</Text> |
| <Text>– Segmentation of a voxel space into free and occupied space: us ∈ [0, 1]</Text> |
| <Text>• Shape prior formulation</Text> |
| <Text>– Voxel space aligned with object of known class</Text> |
| <Text>ix</Text> |
| <Text>s∈ [0, 1] andi</Text> |
| <Text>i x</Text> |
| <Text>sP=1Labeling of a voxel space into 3 labels:</Text> |
| <Text>free space, ground, object</Text> |
| <Text>• Convex Energy</Text> |
| <Text>– Unary term</Text> |
| <Text>∗ Computed from depth maps, local preference for solid class</Text> |
| <Text>– Smoothness term</Text> |
| <Text>∗ Dependent on surface orientation, position and involved labels</Text> |
| </Panel> |
|
|
| <Panel left="495" right="132" width="465" height="378"> |
| <Text>Overview</Text> |
| <Text>• Locally, surface normals similar between different examples</Text> |
| <Text>– Roof at the top of the car close to horizontal</Text> |
| <Figure left="528" right="240" width="406" height="76" no="4" OriWidth="0" OriHeight="0 |
| " /> |
| <Text>• Local distribution of normals captured from training data</Text> |
| <Text>• Input data regularized using trained local normal distributions</Text> |
| <Text>• Trained anisotropic smoothness used for</Text> |
| <Text>– free space ↔ object</Text> |
| <Text>– ground ↔ object</Text> |
| <Text>• ground ↔ free space generic smoothness</Text> |
| <Text>• Label determined by smoothness</Text> |
| </Panel> |
|
|
| <Panel left="495" right="522" width="465" height="331"> |
| <Text>Convex Energy</Text> |
| <Text>∈ </Text> |
| <Text>•iρ</Text> |
| <Text>s :≥</Text> |
| <Text>joint unary term at voxel s for label i</Text> |
| <Text>•ijφ</Text> |
| <Text>s :convex smoothness term at voxel s for labels i and j</Text> |
| <Text>•ix</Text> |
| <Text>s∈ [0, 1]: indicating whether label i is chosen at voxel s</Text> |
| <Text>•ijx</Text> |
| <Text>s−jix</Text> |
| <Text>s3∈ [−1, 1] : represents the local surface orientation</Text> |
| <Text>3• ek ∈ R : k-th canonical basis vector</Text> |
| <Text>• Optimized using primal-dual algorithm [Chambolle and Pock 2011]</Text> |
| </Panel> |
|
|
| <Panel left="973" right="129" width="464" height="205"> |
| <Text>Unary Term</Text> |
| <Text>• Only indicates free or occupied space</Text> |
| <Figure left="1094" right="226" width="280" height="94" no="5" OriWidth="0.315029" OriHeight="0.0764075 |
| " /> |
| </Panel> |
|
|
| <Panel left="973" right="347" width="465" height="253"> |
| <Text>Shape Prior Training</Text> |
| <Figure left="987" right="381" width="456" height="57" no="6" OriWidth="0" OriHeight="0 |
| " /> |
| <Text>• Training data, mesh models</Text> |
| <Text>• Transformed into volumetric models</Text> |
| <Text>• Per voxel s</Text> |
| <Text>– Acquire normal directions of all training samples</Text> |
| <Text>– Generate histogram over normal directions</Text> |
| <Text>– Probability of normal n at s, Ps (n) given by histogram</Text> |
| </Panel> |
|
|
| <Panel left="973" right="614" width="467" height="239"> |
| <Text>Discrete Wulff Shape</Text> |
| <Text>• φs (·) support function of a Wulff shape Wφs</Text> |
| <Text>[Esedoglu and Osher 2004]</Text> |
| <Text>– Wulff shape: convex shape</Text> |
| <Text>• Intersection of half spaces as parameterization of Wφs</Text> |
| <Text>– n half space normal</Text> |
| <Text>–nd</Text> |
| <Text>sdistance of half-space boundary to origin</Text> |
| <Text>• We have φs (n) =nd</Text> |
| <Text>s [Esedoglu and Osher 2004]</Text> |
| <Text>•nd</Text> |
| <Text>s= − log (Ps (n)), determined by training data</Text> |
| <Figure left="1323" right="684" width="119" height="128" no="7" OriWidth="0" OriHeight="0 |
| " /> |
| </Panel> |
|
|
| <Panel left="1450" right="130" width="464" height="155"> |
| <Text>Trained Shape Prior</Text> |
| <Figure left="1468" right="166" width="231" height="98" no="8" OriWidth="0.378613" OriHeight="0.123324 |
| " /> |
| <Figure left="1720" right="168" width="198" height="94" no="9" OriWidth="0.379191" OriHeight="0.142985 |
| " /> |
| <Text>Slices through the bottle shape prior: vertical, horizontal</Text> |
| </Panel> |
|
|
| <Panel left="1449" right="298" width="465" height="455"> |
| <Text>Results</Text> |
| <Figure left="1460" right="338" width="456" height="136" no="10" OriWidth="0.767052" OriHeight="0.195264 |
| " /> |
| <Figure left="1468" right="484" width="447" height="254" no="11" OriWidth="0.772254" OriHeight="0.340036 |
| " /> |
| <Text>Input image</Text> |
| <Text>Depth map</Text> |
| <Text>Vol. fusion</Text> |
| <Text>Shape Prior</Text> |
| </Panel> |
|
|
| <Panel left="1452" right="770" width="463" height="82"> |
| <Text>Acknowledgements</Text> |
| <Text>We gratefully acknowledge the support of the 4DVideo</Text> |
| <Text>ERC starting grant #210806 and V-Charge grant</Text> |
| <Text>#269916 both under the EC’s FP7/2007-2013.</Text> |
| <Figure left="1786" right="813" width="130" height="39" no="12" OriWidth="0" OriHeight="0 |
| " /> |
| </Panel> |
|
|
| </Poster> |
|
|
