Research map / Point clouds / Segmentation and detection
2D-3D Interlaced Transformer for Point Cloud Segmentation with Scene-Level Supervision
Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, Yen-Yu Lin
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 2D multi-view images through interlaced 2D-3D decoder (ablation) | MIT · ScanNet val | mIoU: 31.6 -> 35.8 | Improved |
| add multi-class token transformer encoder (ablation) | 3D CAM baseline · ScanNet val | mIoU: 26.1 -> 31.6 | Improved |
| use interlaced decoder instead of standard cross-attention decoder (ablation) | MIT · ScanNet val | mIoU: 33.7 -> 35.4 | Improved |
| add contrastive N-pair loss on class tokens (ablation) | MIT · ScanNet val | mIoU: 35.4 -> 35.8 | Improved |
| swap query and key-value roles (2D query first) (ablation) | MIT interlaced decoder · ScanNet val | mIoU: 35.4 -> 35.2 | No clear effect |
| use interlaced decoder rather than MLP or BPM fusion (ablation) | MIT · ScanNet val | mIoU: 32.6 (MLP) / 32.4 (BPM) -> 35.8 | Improved |
| add camera poses and depth via 3D world coordinate embedding | MIT · ScanNet val | mIoU: 35.8 -> 37.1 | Improved |
| use BPM fusion with pose and depth instead of MLP fusion (ablation) | MIT · ScanNet val | mIoU: 32.6 -> 32.4 | No clear effect |
| increase number of views beyond 16 (ablation) | MIT · ScanNet pseudo labels | mIoU: 32.7 (16 views) -> 30.9 (32), 31.2 (64) | No clear effect |
| stack more than two interlaced blocks (ablation) | MIT · ScanNet pseudo labels | mIoU: 32.7 (R=2) -> 32.1 (R=3), 32.4 (R=4) | No clear effect |
| remove ImageNet pre-training for 2D ResNet-50 (ablation) | MIT · ScanNet val | mIoU: 35.8 -> 34.6 | Hurt |
| flatten image features into patch tokens instead of global average pooling (ablation) | MIT · ScanNet train (joint 2D-3D) | mIoU: 0.275 (3D) and 0.129 (2D) | Hurt |
Closest papers by meaning
- An MIL-Derived Transformer for Weakly Supervised Point Cloud Segmentation (CVPR 2022 · similarity 0.89)
- 3D Spatial Recognition Without Spatially Labeled 3D (CVPR 2021 · similarity 0.89)
- Dual Adaptive Transformations for Weakly Supervised Point Cloud Segmentation (ECCV 2022 · similarity 0.89)
- WeakPCSOD: Overcoming the Bias of Box Annotations for Weakly Supervised Point Cloud Salient Object Detection (AAAI 2024 · similarity 0.88)
- Multi-Modality Affinity Inference for Weakly Supervised 3D Semantic Segmentation (AAAI 2024 · similarity 0.87)
- Frozen CLIP Transformer Is an Efficient Point Cloud Encoder (AAAI 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.
