Rank2Count

The official repository for the paper "Learning-to-Count by Learning-to-Rank", accepted as an oral presentation at CRV2023. We attempt to solve the problem of weakly supervised object counting using pairwise image ranking.

Motivation

Fully supervised object counting methods typically rely on density map annotations, which are labor intensive to collect. We propose a novel method to exploit pairwise image ranking, which is a significantly weaker form of annotations. These annotations require an annotator to estimate a boolean label $r_{ij} = c_i > c_j$ for an image pair $(x_i, x_j)$, where $c_i$ and $c_j$ represent the true but unknown object count for each image.
In addition to learning directly from pairwise image annotations, we introduces a novel adversarial regularization loss, which encourages the network output to have the structure of a density map while also solving the pairwise ranking problem.

Repository layout

|- src 
|- scripts
|- train.py

Installation

Code

Datasets

Penguins: https://www.robots.ox.ac.uk/~vgg/data/penguins/
Trancos: https://gram.web.uah.es/data/datasets/trancos/index.html
Mall: https://personal.ie.cuhk.edu.hk/~ccloy/downloads_mall_dataset.html

Data format