Group counting is a difficult task because of to extreme occlusion, significant-scale variation, and uneven distribution of people today. Recently, convolutional neural networks have given hope to fix this problem more conveniently. Nevertheless, present-day strategies involve heaps of assorted labeled information in the education process.

Picture credit history: Ben_Kerckx via Pixabay, CC0 Community Area

Hence, a current paper indicates a technique that lessens overfitting and the need to have for expensive labeled information. It makes use of self-supervised transfer colorization learning. Colorization is creating a coloration model of a grayscale photograph. The scientists use the thought that the semantics and neighborhood texture designs attained in the coloration process mirror the density of people today in the region. The experiments on several datasets reveal that the proposed technique achieves much better overall performance given the exact labeled dataset as in comparison with state-of-the-art unlabeled solutions.

Labeled crowd scene photos are high priced and scarce. To appreciably decrease the requirement of the labeled photos, we propose ColorCount, a novel CNN-primarily based approach by combining self-supervised transfer colorization learning and global prior classification to leverage the abundantly out there unlabeled information. The self-supervised colorization department learns the semantics and surface texture of the impression by making use of its coloration factors as pseudo labels. The classification department extracts global group priors by learning correlations amongst impression clusters. Their fused resultant discriminative attributes (global priors, semantics and textures) supply enough priors for counting, for this reason appreciably decreasing the requirement of labeled photos. We conduct comprehensive experiments on 4 demanding benchmarks. ColorCount achieves much much better overall performance as in comparison with other unsupervised strategies. Its overall performance is shut to the supervised baseline with significantly fewer labeled information (ten% of the primary 1).

Investigation paper: Bai, H., Wen, S., and Chan, S.-H. G., “Crowd Counting by Self-supervised Transfer Colorization Mastering and World-wide Prior Classification”, 2021. Website link: muscles/2105.09684