Supplemental Material : Discovering Groups of People in Images

In Fig. 1 and 2, we show additional qualitative examples obtained using our model with poselet [1] and ground truth (GT) detections, respectively. We show the image configuration of groups on the left and corresponding 3D configuration on the right. Different colors and different line types (solid or dashed) represent different groups, the type of each structured group is overlayed on the bottom-left of one participant. In 3D visualization, squares represent standing people, circles represent people sitting on an object, and triangles represent people sitting on the ground. The view point of each individual is shown with a line. The gray triangle is the camera position. The poses are obtained by using the individual pose classification output for visualization purposes. The figures show that our algorithm is capable of correctly associating individuals into multiple different groups while estimating the type of each group. Notice that our algorithm can successfully segment different instances of the same group type that appear in proximity. A distance-based clustering method would not be able to differentiate them. The last figure shows a typical failure case due to only reasoning about people while ignoring objects (such as the tables). Also, we notice that our algorithm can associate individuals into correct groups even in highly complicated scene when GT detections are available.