Adaptive Pattern-Parameter Matching for Robust Pedestrian Detection

Pedestrians with challenging patterns, e.g. small scale or heavy occlusion, appear frequently in practical applications like autonomous driving, which remains tremendous obstacle to higher robustness of detectors. Although plenty previous works have been dedicated these problems, properly matching patterns pedestrian and parameters detector, i.e., constructing a detector proper parameter sizes for certain different complexity, has seldom investigated intensively. Pedestrian instances are usually handled equally the same amount parameters, our opinion is inadequate those more difficult leads unsatisfactory performance. Thus, we propose this paper novel detection approach via adaptive pattern-parameter matching. The input especially complex ones, first disentangled into simpler head by Pattern Disentangling Module (PDM) various receptive fields. Then, Gating Feature Filtering (GFFM) dynamically decides spatial positions where still not simple enough need further disentanglement next-level PDM. Cooperating two key components, can adaptively select best matched size according their complexity. Moreover, explore relationship between performance on corresponding selection policies designed: 1) extending maximum, aiming at occlusion types; 2) specializing group division, variations. Extensive experiments popular benchmarks, Caltech CityPersons, show that proposed method achieves superior compared other state-of-the-art methods subsets scales types.