Abstract
Crowd image is arguably one of the most laborious data to annotate. In this paper, we devote to reduce the massive demand of densely labeled crowd data, and propose a novel weakly-supervised setting, in which we leverage the binary ranking of two images with high-contrast crowd counts as training guidance. To enable training under this new setting, we convert the crowd count regression problem to a ranking potential prediction problem. In particular, we tailor a Siamese Ranking Network that predicts the potential scores of two images indicating the ordering of the counts. Hence, the ultimate goal is to assign appropriate potentials for all the crowd images to ensure their orderings obey the ranking labels. On the other hand, potentials reveal the relative crowd sizes but cannot yield an exact crowd count. We resolve this problem by introducing "anchors" during the inference stage. Concretely, anchors are a few images with count labels used for referencing the corresponding counts from potential scores by a simple linear mapping function. We conduct extensive experiments to study various combinations of supervision, and we show that the proposed method outperforms existing weakly-supervised methods without additional labeling effort by a large margin.
Introduction
The primary hurdle in crowd counting is the "annotation bottleneck." Current state-of-the-art models rely on full supervision, requiring annotators to manually place a dot on every single head in an image. In dense scenes, this can mean thousands of labels per image, a process that is both tedious and prone to error. While some weakly-supervised methods exist—using only the total count per image—even these require exact counts for every training image.
This paper introduces a "glance-to-count" approach based on the observation that humans are significantly better at comparing two images at a glance than providing an exact count. By leveraging simple binary rankings (e.g., "Image A has more people than Image B") and a small set of "anchor" images with known counts, the authors achieve state-of-the-art weakly-supervised performance.
Proposed Framework
The proposed framework shifts the paradigm from direct count regression to ranking potential prediction. It consists of three main stages: labeling, training, and inference.
Siamese Ranking Network (SRN)
The core architecture uses a Siamese network with a CSRNet backbone. Instead of predicting a count, the network predicts a "potential score" ($v$). During training, the model is fed pairs of images. A Ranking Hinge Loss encourages the potential score for the more crowded image to be higher than the other by at least a specific margin ($M$).
Transitive Labeling
To minimize manual effort, the system uses a Directed Acyclic Graph (DAG) to automatically infer rankings. If a human labels $A > B$ and $B > C$, the system automatically detects the transitive relationship $A > C$, significantly reducing the required number of manual comparisons.
The Anchoring Mechanism
Because potential scores are relative, they do not inherently represent absolute counts. To bridge this gap, the authors introduce "anchors"—a tiny subset of images with exact count labels. During inference, a simple linear mapping function (derived via linear regression on the anchor set) translates the predicted potential scores into estimated crowd counts.
Experimental Results
The researchers evaluated their method on major benchmarks including ShanghaiTech Part A, UCF-QNRF, and JHU-CROWD++. The proposed method, particularly the Ours++ (Fully) version which combines ranking labels with available count labels from the anchor set, achieved impressive results.
On ShanghaiTech Part A, the model achieved a Mean Absolute Error (MAE) of 76.9, significantly outperforming other weakly-supervised methods like CCLS (104.6) and approaching the performance of fully supervised models. On the large-scale UCF-QNRF dataset, it set a new benchmark for weakly-supervised performance with an MAE of 133.0. Crucially, the authors reduced the manual labeling effort to roughly 1/4 of the original point-based labels while maintaining high accuracy.
Conclusion
By reframing crowd counting as a ranking problem supplemented by a small set of anchors, this work demonstrates that high-quality models can be trained without the need for exhaustive point-level annotations. The "glance-to-count" approach leverages human intuition for relative density, providing a robust and efficient alternative to traditional supervision methods.
Summary updated