119. LOOKING AT PEDESTRIANS AT DIFFERENT SCALES: A MULTI-RESOLUTION APPROACH AND EVALUATIONS
Department: Electrical & Computer Engineering
Research Institute Affiliation: California Institute for Telecommunications and Information Technology (Calit2)
Faculty Advisor(s): Mohan M. Trivedi
Name: NFN Rakesh Nattoji Rajaram
Grad Year: 2016
Typically, in a detector framework, the model size is fixed and in order to detect pedestrians larger than the model size, the image needs to be down-sampled, leading to information loss which could be vital in detecting larger pedestrians. To this end, we evaluate a multi-resolution detector framework for improved detection of larger pedestrians. In order to demonstrate the effectiveness of the multi-resolution framework, we train the state-of-the-art ACF detector with multiple models in different sizes and analyze its performance. It is to be noted that although this framework is demonstrated to work with ACF detector, almost any detector can be easily integrated into our framework. Our comprehensive evaluation shows that meaningful improvement in detector accuracy can be achieved. Under moderate difficulty settings, on KITTI datasets, we achieve more than 6% increase in detector average precision over the baseline single resolution ACF result. Further insights into improvements in the detector are provided using fine-grained analysis of the detector's performance at various threshold settings.
Industry Application Area(s)