Research map / 3D vision / 3D generation
Malleable 2.5D Convolution: Learning Receptive Fields along the Depth-axis for RGB-D Scene Parsing
Yajie Xing, Jingbo Wang, Gang Zeng
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 |
|---|---|---|---|
| make a_k and t learnable instead of fixed (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.33 -> 48.62 | Improved |
| add learnable kernel rebalancing weights b_k (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.62 -> 48.80 | Improved |
| make only the kernel center a_k learnable (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.28 -> 48.33 | No clear effect |
| learn the depth receptive field versus fixing it (fixed vs full) (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.28 -> 48.80 | Improved |
| increase kernels from 1 to 3 (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 47.08 -> 48.80 | Improved |
| increase kernels from 3 to 5 (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.80 -> 48.17 | Hurt |
| initialize a_k with larger values [-4,-2,0,2,4] (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.80 -> 48.66 | Hurt |
| initialize a_k with smaller values [-1,-0.5,0,0.5,1] (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.80 -> 48.42 | Hurt |
| initialize t to 0.5 instead of 1 (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.80 -> 48.69 | Hurt |
| initialize t to 2 instead of 1 (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.80 -> 48.74 | Hurt |
| use malleable 2.5D conv in all of res2-res5 rather than res2,res3 only (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.27 -> 48.80 | Improved |
| use malleable 2.5D conv in res2-res5 rather than res3-res5 (ablation) | Malleable 2.5D (3 kernels, ResNet-50) · NYUDv2 | mIoU (%): 48.58 -> 48.80 | Improved |
Closest papers by meaning
- Depth-aware CNN for RGB-D Segmentation (ECCV 2018 · similarity 0.90)
- Depth CNNs for RGB-D Scene Recognition: Learning from Scratch Better than Transferring from RGB-CNNs (AAAI 2017 · similarity 0.90)
- DFormerv2: Geometry Self-Attention for RGBD Semantic Segmentation (CVPR 2025 · similarity 0.87)
- DFormer: Rethinking RGBD Representation Learning for Semantic Segmentation (ICLR 2024 · similarity 0.87)
- Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling (CVPR 2024 · similarity 0.86)
- Large Spatial Model: End-to-end Unposed Images to Semantic 3D (NeurIPS 2024 · similarity 0.86)
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.
