Density map estimation enables accurate object counting in heavily occluded, and densely packed scenes where detection-based counting fails. In multi-class density estimation, class awareness can be introduced by modelling classes non-exclusively, better reflecting crowded and visually ambiguous contexts.
This repository contains the code used in our work.
Note
This work was supported by the UKRI AI Centre for Doctoral Training in Sustainable Understandable agri-food Systems Transformed by Artificial INtelligence (SUSTAIN) [grant reference: EP/Y03063X/1].
Published as part of the International Joint Conference on Computational Intelligence at IEEE World Congress of Computational Intelligence 2026.
The paper is available on arXiv, and on the University of Lincoln Figshare Site.
-
Hicks Biodiversity Dataset. Gathered for biodiversity object detection, by Damien Hicks et al. in the Ecological Solutions and Evidence 2021 article Deep learning object detection to estimate the nectar sugar mass of flowering vegetation.
-
VisDrone, a drone based detection challenge for crowded urban areas. We use VisDrone-DET set, so to re-create our results, the trainset, valset and testset-dev must be downloaded. Their GitHub provides links for both Google Drive and BaiduYun downloads.
-
iSAID, a remote sensing segmentation dataset using the images of DOTA-V1.0.
You must download the datasets, and investigate the scripts within convert_datasets. The scripts are seeded, so will produce the same datasets used in this paper. Due to no suitable test/challenge set for iSAID and Hicks et al., those datasets are shuffled to a custom split. Ensure no seeds are changed.
The weights for the backbone must be downloaded from the Twins GitHub Repo.
For now, please email the first author for model weights, although these will eventually be uploaded somewhere open and permanent.