Statistical Optimization of FPN Hyperparameters for improved Pedestrian Detection
Abhishek Gagneja, Amit Kumar Gupta, Brejesh Lall · 2022
Feature Pyramid Network (FPN) has recently been extensively applied in the area of object detection for its multi-scale framework. A majority of tasks utilizing FPN use the pre-trained models or the popular architectural variations as-is. We argue that the architecture to be used depends upon the problem being tackled and that the performance is tightly correlated with the key architecture parameters. We demonstrate our hypothesis on the task of visible pedestrian detection where the bounding boxes are of varied shapes and sizes. To study the impact of various network parameters and to identify suitable values for them, we propose a statistical method based on the property of the dataset being used. We demonstrate the effectiveness of the statistical selection of the parameters on the task of Caltech-USA pedestrian detection and show that with RetinaNet, the modified network performs significantly better on miss rate and mean Average Precision than the baseline network.