| <Poster Width="686" Height="1033"> |
| <Panel left="24" right="158" width="314" height="204"> |
| <Text>SUMMARY!</Text> |
| <Text>This work introduces a framework for recognizing human actions by</Text> |
| <Text>incorporating a new set of visual cues that represent the context of the</Text> |
| <Text>action:!</Text> |
| <Figure left="45" right="224" width="269" height="135" no="1" OriWidth="0.534104" OriHeight="0.204647 |
| " /> |
| </Panel> |
|
|
| <Panel left="24" right="362" width="315" height="58"> |
| <Text>CONTRIBUTIONS!</Text> |
| <Text>Weak foreground-background segmentation approach.
</Text> |
| <Text>• Study of the global camera motion as a cue for action recognition.</Text> |
| <Text>Incorporating appearance from static background.</Text> |
| </Panel> |
|
|
| <Panel left="23" right="422" width="317" height="553"> |
| <Text>METHODOLOGY!</Text> |
| <Text>This work follows the conventional action recognition pipeline. Given a set of</Text> |
| <Text>labeled videos, a set of features is extracted from each video, represented</Text> |
| <Text>using visual descriptors, and combined into a single video descriptor used to</Text> |
| <Text>train a multi-class classifier for action recognition.!</Text> |
| <Figure left="26" right="495" width="307" height="116" no="2" OriWidth="0.53815" OriHeight="0.15773 |
| " /> |
| <Text>• Foreground-background separation: Assuming that a background</Text> |
| <Text>trajectory produces a small frame-to-frame displacement, we associate a</Text> |
| <Text>trajectory with the background if the overall displacement is more than three</Text> |
| <Text>pixels.!</Text> |
| <Figure left="36" right="667" width="292" height="106" no="3" OriWidth="0.53526" OriHeight="0.147453 |
| " /> |
| <Text>• Global camera motion: We argue and show that the relationship between</Text> |
| <Text>an estimated camera motion and underlying action can be a useful cue for</Text> |
| <Text>discriminating certain action classes. As illustrated in the figure below, there</Text> |
| <Text>is a correlation between how the camera moves and the actor.!</Text> |
| <Figure left="53" right="836" width="252" height="135" no="4" OriWidth="0.534104" OriHeight="0.220286 |
| " /> |
| </Panel> |
|
|
| <Panel left="346" right="159" width="316" height="299"> |
| <Text>• Background-context appearance: Beyond local motion and appearance</Text> |
| <Text>properties, the surrounding in which an action is performed is a critical</Text> |
| <Text>component to recognize actions. As Figure below illustrates, the background</Text> |
| <Text>appearance plays an important role to discriminate the action Drumming in</Text> |
| <Text>the sense that the drummer needs a drum set to perform the action.!</Text> |
| <Figure left="375" right="232" width="256" height="117" no="5" OriWidth="0.531792" OriHeight="0.18588 |
| " /> |
| <Text>• Implementation details: We follow two different Bag Of Feature</Text> |
| <Text>implementations as described in the Table below.!</Text> |
| <Figure left="363" right="388" width="275" height="75" no="6" OriWidth="0.42948" OriHeight="0.0437891 |
| " /> |
| </Panel> |
|
|
| <Panel left="350" right="459" width="309" height="325"> |
| <Text>EXPERIMENTAL RESULTS!</Text> |
| <Text>• Datasets: We use state-of-the-art human action datasets and their</Text> |
| <Text>corresponding protocols.!</Text> |
| <Text>• Impact of contextual features: We note that using Fisher vectors</Text> |
| <Text>consistently boost the performance of our contextual features. Also, our</Text> |
| <Text>experiments provide evidence that action recognition performance can be</Text> |
| <Text>improved when static background appearance and global camera motion is</Text> |
| <Text>incorporated with foreground features.!</Text> |
| <Text>Comparison with the state-of-the-art: We </Text> |
| <Text> set </Text> |
| <Text> side </Text> |
| <Text> by </Text> |
| <Text> side </Text> |
| <Text> our </Text> |
| <Text> method </Text> |
| <Text> with </Text> |
| <Text> recent </Text> |
| <Text> methods </Text> |
| <Text> that </Text> |
| <Text> address </Text> |
| <Text> the </Text> |
| <Text> same </Text> |
| <Text> applica5on </Text> |
| <Text> using </Text> |
| <Text> similar </Text> |
| <Text> representa5ons, </Text> |
| <Text> i.e. </Text> |
| <Text> methods </Text> |
| <Text> that </Text> |
| <Text> use </Text> |
| <Text> dense </Text> |
| <Text> trajectory </Text> |
| <Text> points </Text> |
| <Text> to </Text> |
| <Text> represent </Text> |
| <Text> video </Text> |
| <Text> sequences </Text> |
| <Text> [2,3,4] </Text> |
| <Text> in </Text> |
| <Text> the </Text> |
| <Text> Table </Text> |
| <Text> below. </Text> |
| <Text> </Text> |
| <Text> </Text> |
| <Figure left="351" right="647" width="302" height="135" no="7" OriWidth="0.479769" OriHeight="0.122431 |
| " /> |
| </Panel> |
|
|
| <Panel left="347" right="789" width="314" height="187"> |
| <Text>DISCUSSIONS!</Text> |
| <Text>• Contextual features: When combined with foreground trajectories, we show</Text> |
| <Text>that these features, can improve state-of-the-art recognition on challenging</Text> |
| <Text>action datasets.!</Text> |
| <Text>• Project page: http://www.cabaf.net/actioncue!</Text> |
| <Text>References:!</Text> |
| <Text>[1] Fabian Caba Heilbron, Ali Thabet, Juan Carlos Niebles, Bernard Ghanem. Camera Motion</Text> |
| <Text>and Surrounding Scene Appearance as Context for Action Recognition. ACCV, Singapore</Text> |
| <Text>2014.!</Text> |
| <Text>[2] Wang, H., Schmid, C. Action recognition with improved trajectories. ICCV, Sydney 2013.!</Text> |
| <Text>[3] Jiang, Y.G., Dai, Q., Xue, X., Liu, W., Ngo, C.W. Trajectory-based modeling of human</Text> |
| <Text>actions with motion reference points. ECCV, 2012.!</Text> |
| <Text>[4] Jain, M., J egou, H., Bouthemy, P.: Better exploiting motion for better action recognition. !</Text> |
| </Panel> |
|
|
| </Poster> |
|
|
