Research map / Label-efficient and robust learning / Recognition and long-tail
Towards Fine-Grained HBOE with Rendered Orientation Set and Laplace Smoothing
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 RMOS, a 150K-image Blender-rendered dataset with uniform orientation coverage, and OEFormer, a multi-branch transformer with local-window self-attention trained with Laplace-smoothed labels and weighted Smooth-L1 loss for fine-grained human body orientation estimation.
Problem
Human body orientation estimation (HBOE) struggles with fine-grained accuracy because existing datasets (TUD, MEBOW) have non-uniform angle distributions, one-hot classification labels ignore angle similarity, and prior architectures ignore body-part-specific cues.
Approach
Combine a uniformly distributed rendered synthetic dataset (RMOS) with Laplace-smoothed 72-bin soft labels trained via weighted Smooth-L1 loss, and a HRFormer-style multi-branch transformer using local-window self-attention to focus on body parts.
Key results
- OEFormer trained on MEBOW+RMOS reaches MAE 8.129 and Acc.(5°) 72.1 on MEBOW, versus MEBOW-Net's 8.393 MAE and 68.6 Acc.(5°) in the original paper's comparison.
- Ours trained on MEBOW alone reports MAE 8.356 on MEBOW, surpassing Joint-Net (10.526) and PedRecNet (9.702) without synthetic data.
- On TUD, Acc.(15°) rises from 66.7 to 72.5 and Acc.(22.5°) from 77.3 to 83.2 with the proposed method (MEBOW-trained), with MAE 12.298 vs 14.191 for MEBOW-Net.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OEFormer + MEBOW instead of MEBOW-Net on MEBOW | HBOE · MEBOW test | Acc.(5°): 68.6 -> 71.1 | Helped |
| Add RMOS training data to MEBOW-Net (ablation) | MEBOW-Net · MEBOW test | Acc.(5°): 68.4 -> 70.8 | Helped |
| Add more RMOS training data, increasing from 20% to 100% (ablation) | MEBOW-Net · MEBOW test | Acc.(5°): 69.9 -> 70.8 | Helped |
| Use the full OEFormer architecture instead of HRFormer | OEFormer · MEBOW test | Acc.(5°): 69.9 -> 70.6 | Helped |
| Use weighted Smooth-L1-loss rather than Wu et al. smoothing (ablation) | OEFormer + MEBOW · MEBOW test | Acc.(5°): 71.0 -> 72.1 | Helped |
| Use cross-entropy loss instead of weighted Smooth-L1-loss (ablation) | OEFormer + MEBOW · MEBOW test | Acc.(5°): 72.1 -> 44.5 | Hurt |
| Use cosine regression loss instead of weighted Smooth-L1-loss (ablation) | OEFormer + MEBOW · MEBOW test | Acc.(5°): 72.1 -> 31.0 | Hurt |
| Upweight rare sample categories in the loss (ablation) | OEFormer + MEBOW · MEBOW test | Acc.(5°): 72.1 -> 65.3 | Hurt |
| Set Laplace kernel σ to 1.0 instead of 2.0 (ablation) | OEFormer + MEBOW · MEBOW test | Acc.(5°): 72.1 -> 71.0 | No clear effect |
Surprising findings
- Adding RMOS to the TUD-evaluated setting slightly lowered Acc.(5°) (39.5 to 36.6) while improving Acc.(15°) and Acc.(22.5°).
- Upsampling rare orientations instead of adding synthetic RMOS data did not yield the expected gains.
Limitations
- Synthetic RMOS data covers 10 models in 48 poses from five viewpoints, limiting diversity relative to real data.
- Improvements over baselines are often small (e.g., 8.129 vs 8.384 MAE), and gains from RMOS on TUD are mixed.
Experimental setup
Datasets: MEBOW RMOS TUD COCO
Benchmarks: MEBOW TUD
Metrics: Accuracy-5° Accuracy-15° Accuracy-22.5° Accuracy-30° Accuracy-45° MAE
Baselines: MEBOW-Net Joint-Net PedRecNet HRNet HRFormer ResNet50 Hara (2017) AKRF-VW (2018) Yu (2019)
Builds on: HRFormer MEBOW Gaussian label smoothing (Wu et al. 2020) label smoothing (Müller et al. 2019) HRNet
Closest papers by meaning
- Orientation Matters: Making 3D Generative Models Orientation-Aligned (NeurIPS 2025 · similarity 0.84)
- Orient Anything: Learning Robust Object Orientation Estimation from Rendering 3D Models (ICML 2025 · similarity 0.84)
- Orient Anything V2: Unifying Orientation and Rotation Understanding (NeurIPS 2025 · similarity 0.83)
- FreeMan: Towards Benchmarking 3D Human Pose Estimation under Real-World Conditions (CVPR 2024 · similarity 0.83)
- Evaluating Multiview Object Consistency in Humans and Image Models (NeurIPS 2024 · similarity 0.83)
- Neural Localizer Fields for Continuous 3D Human Pose and Shape Estimation (NeurIPS 2024 · similarity 0.83)
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.