Research map / Point clouds / Segmentation and detection
Image-to-Point Cloud Feature Back-Projection for Multimodal Training of 3D Semantic Segmentation
Jiawei Han, Matteo Poggi, Li Huan, Changshuo Wang, Kaiqi Liu, Wei Li
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 |
|---|---|---|---|
| train with IPFP back-projected image features (LiDAR-only inference) | PTv3 · SemanticKITTI val | mIoU: 68.6 -> 71.1 | Improved |
| train with IPFP back-projected image features | PTv3 · nuScenes val | mIoU: 80.3 -> 81.4 | Improved |
| train with IPFP back-projected image features | PTv3 · Waymo val | mIoU: 71.2 -> 72.4 | Improved |
| train with IPFP back-projected image features | SPVCNN · SemanticKITTI val | mIoU: 63.8 -> 65.1 | Improved |
| single-branch IPFP instead of dual-branch fusion (PointPainting, 2DPASS) | PTv3 · SemanticKITTI val | training time (s/it): 0.397 -> 0.204 | Improved |
| swap depth source (IP-BASIC, Depth Anything V2, Unidepth, UnidepthV2) (ablation) | IPFP(PTv3) · SemanticKITTI val | mIoU: 70.87 to 71.13 | No clear effect |
| increase number of back-projected images per scene (ablation) | IPFP(PTv3) · nuScenes val | mIoU: 81.0 -> 81.4 (1 -> 3 images) | Improved |
| stochastically discard back-projected image features during training (ablation) | IPFP(PTv3) · SemanticKITTI val | mIoU: 70.6 -> 71.1 | Improved |
| set discard probability too high (gamma 0.7) (ablation) | IPFP(PTv3) · SemanticKITTI val | mIoU: 71.1 -> 70.5 | Hurt |
| too few or too many back-projected cluster centers (2% or 7% vs 5%) (ablation) | IPFP(PTv3) · SemanticKITTI val | mIoU: 71.1 -> 70.4 (2%), 70.8 (7%) | Hurt |
| use pretrained lightweight CNNs instead of learnable conv+linear for cluster module (ablation) | IPFP(PTv3) · nuScenes val | mIoU: 81.4 -> 80.7 / 80.6 | Hurt |
| raise lower depth percentile (alpha_l) to 50 to drop near-field samples (ablation) | IPFP(PTv3) · SemanticKITTI val | mIoU: 70.7 (20) -> 71.1 (50) -> 70.9 (60) | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- EPMF — EPMF: Efficient Perception-Aware Multi-Sensor Fusion for 3D Semantic Segmentation (TPAMI 2024) · better on 1 of 1 shared result
- 2DPASS — 2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds (ECCV 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR-Based Perception (TPAMI 2021 · similarity 0.90)
- Perception-Aware Multi-Sensor Fusion for 3D LiDAR Semantic Segmentation (ICCV 2021 · similarity 0.90)
- Semantic Segmentation for Real Point Cloud Scenes via Bilateral Augmentation and Adaptive Fusion (CVPR 2021 · similarity 0.90)
- Point-MoE: Large-Scale Multi-Dataset Training with Mixture-of-Experts for 3D Semantic Segmentation (ICLR 2026 · similarity 0.89)
- Multimodality Helps Few-shot 3D Point Cloud Semantic Segmentation (ICLR 2025 · similarity 0.88)
- Geometrically-driven Aggregation for Zero-shot 3D Point Cloud Understanding (CVPR 2024 · similarity 0.87)
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.
