Research map / Label-efficient and robust learning / Domain adaptation and generalization

Single Domain Generalization for Crowd Counting

CVPR 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Foundation-model domain generalization · 37 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MPCount, a single domain generalization method for density-map crowd counting that uses one attention memory bank, a content error mask, attention consistency loss, and patch-wise classification to handle regression and label ambiguity under narrow source domains.

Problem

Crowd counting models trained with density map regression degrade under domain shift, and existing single domain generalization methods for classification and segmentation do not transfer well to regression with ambiguous pixel-level labels, especially when the source domain is narrow.

Approach

An attention memory bank reconstructs features as attention over a single memory of density values, with a content error mask filtering domain-related elements, an attention consistency loss aligning attention across original and photometrically augmented features, and a patch-wise classification auxiliary task that masks density predictions in background patches to counter label ambiguity.

Key results

• On SHB→SHA, MPCount reduces counting error by 18.2% relative to the prior state of the art, and on SN→FH by 21.8%.
• Under the A→B setting, MPCount reaches MAE 11.4 and MSE 19.7, the best among tested DG methods.
• On narrow JHU-Crowd++ domains such as SD→SR and SN→FH, MPCount achieves MAE 37.4 and 216.3 respectively, outperforming DCCUS and the normalization-based SDG methods.

Measured effects reported in the paper

Surprising findings

• Normalization/whitening-based SDG methods (IBN, SW, ISW) adapted to counting perform worse than fully supervised counting baselines on narrow-domain settings.
• DCCUS performs worse than fully supervised methods when the source domain distribution is narrow, since sub-domain clustering may fail.

Limitations

• Relies on heuristic hyperparameters (memory size, alpha threshold, loss weights) tuned on the A→B setting.
• Evaluated only on crowd counting benchmarks with VGG16-BN backbone; generalization to other architectures is not reported.

Experimental setup

Datasets: ShanghaiTech Part A ShanghaiTech Part B UCF-QNRF JHU-Crowd++

Benchmarks: ShanghaiTech Part A ShanghaiTech Part B UCF-QNRF JHU-Crowd++

Metrics: MAE MSE

Baselines: BL DM-Count SASNet ChfL MAN RBT C2MoT FGFD DAOT FSIM IBN SW

Builds on: DCCUS instance normalization memory bank for classification

Closest papers by meaning

• Domain-General Crowd Counting in Unseen Scenarios (AAAI 2023 · similarity 0.92)
• Error-Aware Density Isomorphism Reconstruction for Unsupervised Cross-Domain Crowd Counting (AAAI 2021 · similarity 0.88)
• Leveraging Self-Supervision for Cross-Domain Crowd Counting (CVPR 2022 · similarity 0.88)
• Leveraging Unlabeled Data for Crowd Counting by Learning to Rank (CVPR 2018 · similarity 0.88)
• Distribution Matching for Crowd Counting (NeurIPS 2020 · similarity 0.88)
• Counting without Scale: Scale-Consistent Error Correction for Crowd Counting (NeurIPS 2026 · similarity 0.88)