H2P×PKD: Progressive Training Pipeline with Knowledge Distillation for Lightweight Backbones in Pedestrian Detection

Pedestrian Detection, an instance of Object Detection, has many applications, especially in autonomous driving and surveillance. While many methods are tailored for pedestrian detection, studies suggest that general object detection approaches can also be effective with the appropriate training settings. The large domain unsimilarity between general object data and pedestrian-specific data is a significant challenge. To address this, Pedestron introduces a progressive training pipeline that begins with pre-training on a large and diverse general human dataset, followed by fine-tuning on pedestrian-specific data. However, this transition remains too abrupt. Additionally, the model’s size and computational cost are also not receiving much attention. Therefore, we propose H2P, a new training pipeline based on Pedestron. The pipeline includes IDSS, an intermediate training step for detecting general humans in traffic-related contexts before fine-tuning on pedestrian datasets. Moreover, we leverage PKD, a knowledge distillation method via Pearson Correlation Coefficient, to improve lightweight models using FPN features from large-scale, general teacher models. To evaluate the effectiveness of our H2P with PKD approach, we experiment on two lightweight backbones: a traditional MobileNetV2 and a downscaled variant of InternImage, named InternImage-M. Our best result is 8.59% MR<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−2</sup> on CityPersons’ validation set. Our lightweight InternImage-M backbone gives the promising result of 11.53% MR<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−2</sup>.