ValueUp-P2PNet: Suppressing Duplicate Proposals in Point-Based Crowd Counting via an Assignment-Aware Uniqueness Branch

Crowd counting estimates the number of people in an image while reliably representing individual locations. Density map-based methods are robust in dense scenes but cannot directly output individual locations, whereas the point-based P2PNet predicts head locations as point proposals and learns one-to-one assignments with ground-truth (GT) points via Hungarian matching. However, our proposal-level analysis reveals a duplicate proposal problem: multiple proposals can remain around a single GT point even when the counting error is small, so count-level metrics alone cannot evaluate structural prediction quality. The proposed ValueUp-P2PNet preserves the location and confidence prediction structure of P2PNet while predicting a uniqueness score reflecting each proposal's representativeness. Using matching results as supervision, it raises the uniqueness of representative proposals and lowers that of nearby duplicates. A Gaussian-weighted duplicate loss alleviates threshold discontinuity, and an adaptive loss-ratio curriculum adjusts the auxiliary loss contribution with training progress. On ShanghaiTech Part A, UCF-CC-50, and UCF-QNRF, ValueUp-P2PNet achieves comparable or lower MAE than the original P2PNet under the same setting while consistently increasing the matched-proposal ratio (Matched/GT) and reducing the duplicate-proposal ratio (Dup/Pred) on all three datasets. An ablation study shows that each loss term contributes differently to counting performance and to proposal-set structural quality.