Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Imbalanced Regression via Hierarchical Classification Adjustment
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 Hierarchical Classification Adjustment (HCA) for imbalanced regression, combining an ensemble of progressively coarser quantized classifiers and distilling them into a single range-preserving classifier (HCA-d).
Problem
Regression tasks in computer vision (age estimation, crowd counting, depth estimation) suffer from imbalanced target distributions; classification-by-quantization faces a trade-off between class balance/sufficiency and quantization error.
Approach
Train H hierarchical classifiers with progressively coarser, count-balanced quantizations, adjust the finest classifier's logits with coarser ones, and distill the ensemble into one classifier using a max-based range-preserving alignment so predicted ranges stay consistent across the hierarchy.
Key results
- On IMDB-WIKI-DIR, HCA-d+LDS+ranksim reaches overall bMAE 11.92 versus 12.33 for CLS+LDS+ranksim.
- On SHTech Part A, HCA-d reduces MAE to 53.7 from 58.2 for the vanilla classifier and 60.7 for DC-regression.
- Sum-based alignment (eq. 10) degrades HCA to 150.3 overall bMAE, while the range-preserving max alignment gives 53.7, showing range preservation is key.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add range-preserving hierarchical distillation | CLS · IMDB-WIKI-DIR | bMAE (All): 13.58 -> 12.70 | Helped |
| Add range-preserving hierarchical distillation | CLS · NYUDv2-DIR | MAE: 1.011 -> 0.987 | Helped |
| Add range-preserving hierarchical distillation | CLS · SHA | MAE: 58.2 -> 53.7 | Helped |
| Add range-preserving hierarchical distillation | CLS · SHB | MAE: 7.0 -> 6.8 | Helped |
| Add LDS to HCA-d | CLS+LDS baseline · IMDB-WIKI-DIR | bMAE (All): 12.85 -> 12.42 | Helped |
| Add HCA-d to CLS+LDS+ranksim | CLS+LDS+ranksim · IMDB-WIKI-DIR | bMAE (All): 12.33 -> 11.92 | Helped |
| Add HCA-d to CLS+LDS+ranksim | CLS+LDS+ranksim · IMDB-WIKI-DIR many-shot | bMAE: 6.70 -> 6.88 | Hurt |
| Use HCA-d instead of HCA-add (ablation) | Hierarchical classifier ensemble · IMDB-WIKI-DIR | bMAE (All): 12.86 -> 12.70 | Helped |
| Use HCA-d instead of HCA-mul (ablation) | Hierarchical classifier ensemble · SHA | MAE: 54.7 -> 53.7 | Helped |
| Equalize sample counts rather than interval lengths in hierarchical class splitting (ablation) | HCA-d · IMDB-WIKI-DIR | bMAE (All): 12.77 -> 12.70 | Helped |
| Use range-preserving max alignment instead of sum alignment (ablation) | HCA distillation · IMDB-WIKI-DIR | bMAE (All): 27.08 -> 12.70 | Helped |
| Use range-preserving max alignment instead of sum alignment (ablation) | HCA distillation · SHA | MAE: 150.3 -> 53.7 | Helped |
Surprising findings
- Naive summation-based hierarchical alignment harms performance badly, with overall bMAE rising to 150.3 versus 58.2 for the single classifier.
- HCA gives little improvement on NYUDv2-DIR because its classes have sufficient samples, consistent with the theory.
Limitations
- Theoretical analysis is limited to two classifiers and is extended to H by induction only as a sketch.
- Improvements are small on datasets with sufficient samples per class such as NYUDv2-DIR.
Experimental setup
Datasets: IMDB-WIKI-DIR SHTech NYUDv2-DIR
Benchmarks: IMDB-WIKI-DIR SHTech Part A SHTech Part B NYUDv2-DIR
Metrics: bMAE MAE RMSE AbsRel delta accuracy thresholds
Baselines: Regression Regression+LDS Regression+LDS+ranksim Balanced MSE CLS CLS+LDS CLS+LDS+ranksim CSRNet BL MNA OT GL
Builds on: Balanced MSE label distribution smoothing soft ordinal loss (SORD) ranksim hierarchical classification distillation
Closest papers by meaning
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.86)
- Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.85)
- Improve Representation for Imbalanced Regression through Geometric Constraints (CVPR 2025 · similarity 0.85)
- Leveraging Group Classification with Descending Soft Labeling for Deep Imbalanced Regression (AAAI 2025 · similarity 0.85)
- Balanced Data, Imbalanced Spectra: Unveiling Class Disparities with Spectral Imbalance (ICML 2024 · similarity 0.85)
- Learning under Localized Minority Imbalance (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.