Research map / 3D vision / Human pose estimation
Pedestrian-Centric 3D Pre-collision Pose and Shape Estimation from Dashcam Perspective
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.
Builds PVCP, a dashcam pedestrian-vehicle collision pose/shape dataset (42k frames, ~19.5k annotated pre-collision poses with 2D, 3D and SMPL labels), and PPSENet, a two-stage ITP+PTM network using a MotionBERT backbone, iterative regression and a pose-class loss.
Problem
Estimating pedestrian pre-collision pose and body shape from dashcam accident videos is poorly handled by pose models trained on daily-activity datasets, and in-the-wild 3D pose data for collisions is scarce.
Approach
Combine a semi-automatically annotated collision-pose dataset with a two-stage pipeline: a ResNet50 heatmap 2D pose estimator followed by a MotionBERT-initialized DSTformer lifting network with iterative SMPL regression and an auxiliary four-class pose classification loss.
Key results
- On the PVCP testset with 2D detected input, PPSENet reaches MPVE 257.75 mm vs. 282.50 mm for MotionBERT baseline and 849-857 mm for VIBE.
- With 2D ground-truth input, the full model reaches MPVE 140.43 mm and MPJPE_17j 107.47 mm, versus 145.77 mm MPVE for MotionBERT fine-tuned without iterations or class loss.
- Pose-class-aware PPSENet achieves the best PA-MPVE/PA-MPJPE among compared methods on the PVCP testset, per the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pretrain on pose datasets and train on PVCP instead of PVCP training alone (ablation) | PTM with MotionBERT · PVCP testset (All), 2D Det input | MPVE (mm): 315.64 -> 282.50 | Helped |
| pretrain on pose datasets and train on PVCP instead of PVCP training alone (ablation) | PTM with MotionBERT · PVCP testset (All), 2D GT input | MPVE (mm): 300.04 -> 145.77 | Helped |
| add iterative decoder with 3 iterations (ablation) | PTM with MotionBERT · PVCP testset (All), 2D Det input | MPVE (mm): 282.50 -> 266.20 | Helped |
| add iterative decoder with 3 iterations (ablation) | PTM with MotionBERT · PVCP testset (All), 2D GT input | MPVE (mm): 145.77 -> 145.75 | No clear effect |
| add pose class loss (ablation) | PTM with MotionBERT · PVCP testset (All), 2D Det input | MPVE (mm): 266.20 -> 259.05 | Helped |
| add pose class loss (ablation) | PTM with MotionBERT · PVCP testset (All), 2D GT input | MPVE (mm): 145.75 -> 141.28 | Helped |
| use 3 iterations instead of 2 (ablation) | PTM with MotionBERT and class loss · PVCP testset (All), 2D GT input | MPVE (mm): 141.95 -> 140.43 | Helped |
| use 4 iterations instead of 3 (ablation) | PTM with MotionBERT and class loss · PVCP testset (All), 2D GT input | MPVE (mm): 140.43 -> 139.96 | Helped |
| use 5 iterations instead of 3 (ablation) | PTM with MotionBERT and class loss · PVCP testset (All), 2D GT input | MPVE (mm): 140.43 -> 140.01 | Helped |
| use 6 iterations instead of 3 (ablation) | PTM with MotionBERT and class loss · PVCP testset (All), 2D GT input | MPVE (mm): 140.43 -> 140.41 | Helped |
| fine-tune on PVCP rather than train only on COCO (ablation) | PTM · PVCP testset (All), 2D GT input | MPVE (mm): 185.57 -> 140.43 | Helped |
| fine-tune on PVCP rather than train only on PW3D (ablation) | PTM · PVCP testset (All), 2D GT input | MPVE (mm): 196.33 -> 140.43 | Helped |
Surprising findings
- Pretraining alone on large pose datasets performed worse than training on PVCP alone with 2D detected inputs (MPVE 335.11 mm vs 315.64 mm), and models trained on COCO or PW3D without fine-tuning performed poorly on PVCP.
Limitations
- Dataset is small (209 videos) and lacks camera parameters, vehicle speed, and global pedestrian position/direction.
- Two-stage design makes final pose error depend heavily on 2D pose estimation, which degrades on occluded lower limbs (ankle PCK 1.93%).
- Not real-time because input is image plus pre-selected bounding box sequences.
Experimental setup
Datasets: PVCP AMASS Human3.6M PW3D MSCOCO MPII JHMDB PedX
Benchmarks: PVCP testset
Metrics: MPJPE PA-MPJPE MPVE PA-MPVE PCK
Baselines: VIBE PARE Pose2Mesh MotionBERT
Builds on: MotionBERT SPIN Pose2Mesh VIBE PARE
Closest papers by meaning
- PoseHD: Boosting Human Detectors Using Human Pose Information (AAAI 2018 · similarity 0.84)
- 3D Human Pose Estimation Using Spatio-Temporal Networks with Explicit Occlusion Training (AAAI 2020 · similarity 0.84)
- Capturing Humans in Motion: Temporal-Attentive 3D Human Pose and Shape Estimation From Monocular Video (CVPR 2022 · similarity 0.84)
- Attention-Aware Compositional Network for Person Re-Identification (CVPR 2018 · similarity 0.84)
- FreeMan: Towards Benchmarking 3D Human Pose Estimation under Real-World Conditions (CVPR 2024 · similarity 0.83)
- Optimizing Human Pose Estimation Through Focused Human and Joint Regions (AAAI 2025 · 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.