Globally-optimal greedy algorithms for tracking a variable number of objects

Abstract

Problem input: • Detector scores for all space time windows in a video Problem output: • Number of objects (tracks) • Starting and ending frames for each track • Tracks themselves Our contributions: • Globally Optimal (for a common class of objective functions) • Locally Greedy (and hence straightforward to implement) • Scale linearly in the number of objects and (quasi)linearly with video-length A detection window A transition Has input flow of d Has output flow of d Equivalent graph problem: Min-cost flow Solutions: • Globally optimum • Push-relabel algorithm • Used in Zhang et al CVPR'08 • Solves for known d; Does binary search to find K=optimum d • Successive shortest path • Special structure in our graph (DAG, unit-capacity) • Is greedy • Approximate • Dynamic programming • Is greedy Successive shortest path: • At each iteration, find the shortest path and augment the graph by reversing edges along it. • Iterate as long as the cost of shortest path is positive. Dynamic programming: Sweep the graph and update paths to find the shortest one • Variable length DP because of " s " and " t " nodes • Is optimum for since the graph is DAG. • Is not optimum for next iterations • Ignore backward edges: works since the optimum solution rarely uses them. • 2-pass DP: Run one more DP along backward edges • Is not optimum because of very rare cases. • Can do multi-pass DP Caching: 1000X faster • Use shortest paths from previous iterations. Keep track of the origin for each path • After instancing a track, update only paths which share the same origin. • Non-max-suppression in the loop: • At each iteration, suppress all windows overlapping with the instanced track. • Handling occlusion: • Adding more than one-frame jumps to transition edges. • Markov model places geometric prior on track length (short tracks more likely). • Approximate DP algorithm tends to produce longer tracks which may be more accurate

DOI: 10.1109/CVPR.2011.5995604

Extracted Key Phrases

Showing 1-10 of 244 extracted citations
0501002011201220132014201520162017
Citations per Year

359 Citations

Semantic Scholar estimates that this publication has received between 299 and 436 citations based on the available data.

See our FAQ for additional information.