Research map / 3D vision / Human pose estimation
Hidden Footprints: Learning Contextual Walkability from 3D Human Trails
Jin Sun, Hadar Averbuch-Elor, Qianqian Wang, Noah Snavely
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hidden footprint propagation (ablation) | MSE-trained model · Waymo 50-scene manually labeled set | mAP: 0.111 -> 0.145 | Improved |
| add class-balanced (reweighted) BCE instead of standard BCE (ablation) | footprint-trained model · Waymo 50-scene manually labeled set | mAP: 0.119 -> 0.190 | Improved |
| add adversarial feature loss to reweighted BCE (ablation) | RBCE model · Waymo 50-scene manually labeled set | mAP: 0.190 -> 0.269 | Improved |
| add class-balanced loss to adversarial-only training (ablation) | adversarial-only model · Waymo 50-scene manually labeled set | mAP: 0.173 -> 0.269 | Improved |
| add class-balanced + adversarial losses instead of MSE (ablation) | MSE model · Waymo 50-scene manually labeled set | mAP: 0.145 -> 0.269 | Improved |
| add adversarial feature loss to reweighted BCE (ablation) | RBCE model · Waymo entire test set | Pred Valid (TP) ratio: 0.53 -> 0.69 | Improved |
| add class-balanced loss to adversarial-only training (ablation) | adversarial-only model · Waymo entire test set | Pred Total (TP+FP) ratio: 22.15 -> 8.18 | Hurt |
| use focal loss instead of class-balanced loss (ablation) | Ours · Waymo 50-scene manually labeled set | performance | No clear effect |
| propagate 3D person labels across the sequence | Waymo labels · Waymo train | locations per image: 3.8 -> 362 | Improved |
| use proposed walkability model instead of semantic-segmentation-based Semantic-Net | Semantic-Net · Waymo 50-scene manually labeled set | mAP: 0.201 -> 0.269 | Improved |
| use proposed model instead of Inpainting baseline | Inpainting · Waymo 100-scene manually labeled set | mAP: 0.107 -> 0.246 | Improved |
| use Waymo-trained model on Cityscapes instead of Cityscapes-trained model of Lee et al. | Lee et al. · Cityscapes val | KL divergence: 0.78 -> 0.42 | Improved |
Closest papers by meaning
- WalkGPT: Grounded Vision-Language Conversation with Depth-Aware Segmentation for Pedestrian Navigation (CVPR 2026 · similarity 0.85)
- mmWalk: Towards Multi-modal Multi-view Walking Assistance (NeurIPS 2025 · similarity 0.84)
- Learning to Generate Diverse Pedestrian Movements from Web Videos with Noisy Labels (ICLR 2025 · similarity 0.83)
- Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking (NeurIPS 2024 · similarity 0.83)
- Text to Blind Motion (NeurIPS 2024 · similarity 0.83)
- CityWalker: Learning Embodied Urban Navigation from Web-Scale Videos (CVPR 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.
