Active Deformable Part Models

TL;DR

This paper introduces an active, optimized approach for deformable part model-based object detection that improves speed by learning part evaluation order and stopping criteria, maintaining accuracy on standard datasets.

Contribution

It proposes a novel active detection method that optimizes part evaluation order and stopping time using learned statistics and dynamic programming.

Findings

Faster detection than traditional cascade methods

Negligible accuracy loss on PASCAL VOC datasets

Effective part scheduling improves efficiency

Abstract

This paper presents an active approach for part-based object detection, which optimizes the order of part filter evaluations and the time at which to stop and make a prediction. Statistics, describing the part responses, are learned from training data and are used to formalize the part scheduling problem as an offline optimization. Dynamic programming is applied to obtain a policy, which balances the number of part evaluations with the classification accuracy. During inference, the policy is used as a look-up table to choose the part order and the stopping time based on the observed filter responses. The method is faster than cascade detection with deformable part models (which does not optimize the part order) with negligible loss in accuracy when evaluated on the PASCAL VOC 2007 and 2010 datasets.