The effects of image smoothing on CNN-based detectors

Edge-preserving smoothing filters have been shown to improve generalisation performance on the HOG features with a SVM classifier. However, not all smoothing filters and parameters lead to better performance. The effects of smoothing filters are studied on the Faster R-CNN detector using generic object and human detection datasets, namely the PASCAL VOC and KITTI respectively. The total variation (TV) smoothing filter was used for this study. It was found that the TV smoothing removed details the CNN was using for detection which degraded performance for both datasets. The results are consistent with previous observations that CNNs tend to learn weak visual features. The performance loss, however, was moderate and could be justified in the context of improving robustness to perturbations. The PASCAL VOC and KITTI datasets had comparable performance loss despite the latter having many more small objects that tend to blend into the background when smoothing is applied.

I. Introduction

Human detection is a fundamental problem in computer vision with many important applications, such as pedestrian safety, intelligent surveillance and human-machine interfaces. Human detection is one of the many beneficiaries of the progress made on CNNs. The CNN resembles the hierarchy of the nervous system with layer by layer abstraction of the image data to mine information [45]. The relatively small size of human detection datasets initially presented an obstacle to the use of CNNs. Sermanet et al. were the first to overcome this by using unsupervised pre-training [35]. Hosang et al. [15] successfully used the CifarNet model, making them the first to use a vanilla CNN model. Datasets such as the Caltech Pedestrian and KITTI have superseded earlier datasets like INRIA in size and level of annotation [49].