Scale-Invariant Continuous Implicit Neural Representations For Object Counting

Siyuan Xu, Yucheng Wang, Xihaier Luo, Byung-Jun Yoon, Xiaoning Qian

Keywords: scale invariance, implicit neural representation, object counting

TL;DR: A scale-invariant continuous representation learner.

Abstract:

Many object counting methods rely on density map estimation (DME) using convolutional neural networks (CNNs) on discrete grid image representations. However, these methods struggle with large variations in object size or input image resolution, typically due to different imaging conditions and perspective effects. Worse yet, discrete grid representations of density maps result in information loss with blurred or vanished details for low-resolution inputs.
To overcome these limitations, we design Scale-Invariant Implicit neural representations for counting (SI-INR) to map arbitrary-scale input signals into a continuous function space, where each function produces density values over continuous spatial coordinates. SI-INR achieves robust counting performances with respect to changing object sizes, extensive experiments on commonly used diverse datasets have validated the proposed method.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5092