Curriculum for Crowd Counting -- Is it Worthy?

TL;DR

This paper rigorously evaluates curriculum learning (CL) in crowd counting using density estimation across eight models and two ShanghaiTech datasets. By conducting roughly 112 experiments with six pacing strategies and two scoring variants, it shows that CL can substantially improve accuracy in certain settings while offering marginal gains in others, with the pacing function playing a critical role. A consistent finding is that CL can reduce convergence time, even when accuracy gains are modest. The work highlights the conditional nature of CL's benefits and suggests extending CL analyses to other computer vision tasks and scoring mechanisms for clearer guidance on when to adopt CL in supervised learning.

Abstract

Recent advances in deep learning techniques have achieved remarkable performance in several computer vision problems. A notably intuitive technique called Curriculum Learning (CL) has been introduced recently for training deep learning models. Surprisingly, curriculum learning achieves significantly improved results in some tasks but marginal or no improvement in others. Hence, there is still a debate about its adoption as a standard method to train supervised learning models. In this work, we investigate the impact of curriculum learning in crowd counting using the density estimation method. We performed detailed investigations by conducting 112 experiments using six different CL settings using eight different crowd models. Our experiments show that curriculum learning improves the model learning performance and shortens the convergence time.

Figures (4)

• Figure 1: Curriculum learning framework.
• Figure 2: Various pacing functions applied to ShanghaiTech Part B dataset with a total number of samples (N) = 400 and batch size = 8.
• Figure 3: Pacing functions used in our experiments.
• Figure 4: Illustration of model convergence using standard training versus curriculum learning.