| <Poster Width="1734" Height="1226"> |
| <Panel left="20" right="192" width="531" height="990"> |
| <Text>1. Abstract</Text> |
| <Text>• We formulate the problem of joint visual</Text> |
| <Text>attribute and object class image segmentation</Text> |
| <Text>as a dense multi-labelling problem, where each</Text> |
| <Text>pixel in an image can be associated with both</Text> |
| <Text>an object class and a set of visual attributes</Text> |
| <Text>labels.</Text> |
| <Text>• In order to learn the label correlations, we</Text> |
| <Text>adopt a boosting-based piecewise training</Text> |
| <Text>approach with respect to the visual appearance</Text> |
| <Text>and co-occurrence cues.</Text> |
| <Text>Weuseafiltering-basedmean-field</Text> |
| <Text>approximation approach for efficient joint</Text> |
| <Text>inference. Further, we develop a hierarchical</Text> |
| <Text>model to incorporate region-level object and</Text> |
| <Text>attribute information.</Text> |
| <Text>Object class segmentation</Text> |
| <Text>• Assigning an object class label to each pixel</Text> |
| <Figure left="29" right="706" width="511" height="174" no="1" OriWidth="0.379469" OriHeight="0.0748663 |
| " /> |
| <Text>Image Segmentation with Objects and Attributes</Text> |
| <Text>• Assigning an object class label and a set of</Text> |
| <Text>visual attribute labels to each pixel</Text> |
| <Figure left="26" right="987" width="515" height="175" no="2" OriWidth="0.3812" OriHeight="0.0766488 |
| " /> |
| </Panel> |
|
|
| <Panel left="585" right="191" width="515" height="648"> |
| <Figure left="586" right="237" width="516" height="367" no="3" OriWidth="0.369666" OriHeight="0.202317 |
| " /> |
| <Figure left="590" right="648" width="506" height="171" no="4" OriWidth="0" OriHeight="0 |
| " /> |
| </Panel> |
|
|
| <Panel left="586" right="841" width="516" height="337"> |
| <Figure left="591" right="914" width="509" height="247" no="5" OriWidth="0.795848" OriHeight="0.289216 |
| " /> |
| <Text> Fully-connected CRF</Text> |
| <Text> Joint Pixel-level CRF</Text> |
| <Text> Hierarchical CRF</Text> |
| <Text> Ground Truth</Text> |
| </Panel> |
|
|
| <Panel left="1150" right="194" width="517" height="646"> |
| <Text>4. Attribute-augmented NYU dataset</Text> |
| <Figure left="1153" right="242" width="515" height="313" no="6" OriWidth="0.784314" OriHeight="0.3988425 |
| " /> |
| <Text> Attribute annotation</Text> |
| <Text> Image</Text> |
| <Text> Object annotation</Text> |
| <Text>Following the CORE dataset, we augment the</Text> |
| <Text>attribute annotations for NYU V2 dataset. Above</Text> |
| <Text>figure shows the annotations for aNYU dataset.</Text> |
| <Text>Below figure demonstrate the annotation of CORE</Text> |
| <Text>dataset and aPASCAL dataset.</Text> |
| </Panel> |
|
|
| <Panel left="1144" right="842" width="569" height="321"> |
| <Text>5. Acknowledgement</Text> |
| <Text>This project is supported by EPSRC EP/I001107/2,</Text> |
| <Text>ERC HELIOS 2013-2018Advanced Investigator Award.</Text> |
| <Text>[1] Dense semantic image segmentation</Text> |
| <Text>with objects and attributes. CVPR, 2014.</Text> |
| <Text>[2] ImageSpirit: Verbal Guided Image</Text> |
| <Text>Parsing, ACM TOG 2014.</Text> |
| <Text>[3] Efficient Inference in Fully Connected</Text> |
| <Text>CRFs with Gaussian Edge Potentials.</Text> |
| <Text>NIPS 2011.</Text> |
| <Text>http://kylezheng.org/densesegattobj/</Text> |
| <Figure left="1539" right="961" width="172" height="173" no="7" OriWidth="0" OriHeight="0 |
| " /> |
| </Panel> |
|
|
| </Poster> |
|
|
