Marked Point Processes for Crowd Counting Weina Ge and Robert T. Collins Computer Science and Engineering Department, The Pennsylvania State University, USA Introduction Extrinsic Shape Mappings Estimation Experimental Results Detection Results A Bayesian marked point process (MPP) model is developed to detect and count people in crowded scenes. The model couples a spatial stochastic process governing number and placement of individuals with a conditional mark process for selecting body shape. We automatically learn the mark (shape) process from training video by estimating a mixture of Bernoulli shape prototypes along with an extrinsic shape distribution describing the orientation and scaling of these shapes for any given image location. The reversible jump Markov Chain Monte Carlo framework is used to efficiently search for the maximum a posteriori configuration of shapes, leading to an estimate of the count, location and pose of each person in the scene. Intuition: consider a center-surround region of a given scale, centered at a given pixel We have evaluated CSDD performance with respect to detection repeatability and matching utility. Details of the experiments and a complete set of results can be found on our website: 1. Extract feature distribution F 2. Extract feature distribution G http://vision.cse.psu.edu/projects/mpp/mpp.html 3. Compute Earth Mover’s Distance EMD(F,G) to measure dissimilarity of center region from surround region. Matching Results How to do this efficiently for all pixels? binary channels channels We use a marked point process to determine the number and configuration of multiple people in a scene. In addition to determining the location, scale and orientation of each individual, the MPP also selects an appropriate body shape from a set of learned Bernoulli shape prototypes, as displayed at the bottom. CSDD score (EMD) Motivation Implementation Details Soft map This method of EMD computation only works for 1D distributions. For n-D distributions, we concatenate the n 1D marginals to get a 1D distribution. Fast LoG filtering at every scale is performed using a fourth-order IIR filter (aka Deriche-filtering). We form a scale space of CSDD score images indexed by the scale of the LoG filter. CSDD features are then found as extrema in both scale and space. Row 1: four frames from a parking lot video sequence, showing affine alignment of bottom frame overlaid on top frame. Row 2: left to right: shout3 to shout4; shout2 to was2 (images courtesy of Tinne Tuytelaars); stop sign; snowy stop sign. Row 3: kampa1 to kampa4 (images courtesy of Jiri Matas); bike1 to bike6; trees1 to trees5; ubc1 to ubc6. Row 4: natural textures: asphalt; grass; gravel; stones. Example: Yin-yang symbol superimposed on an intensity gradient. Of the six interest region detectors compared, only the CSDD detector captures the natural location and scale of the symbol.