Research map / Architectures and efficiency / CNN design
SMPConv: Self-Moving Point Representations for Continuous Convolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace FlexTCN's parameterized kernels with self-moving point kernels | FlexTCN · sMNIST and pMNIST | accuracy: 99.62 -> 99.75; 98.63 -> 99.10 | Helped |
| replace FlexTCN's parameterized kernels with self-moving point kernels | FlexTCN · sCIFAR10 | accuracy: 80.82 -> 84.86 | Helped |
| replace FlexTCN's parameterized kernels with self-moving point kernels | FlexTCN · SC-raw | accuracy: 91.73 -> 94.95 | Helped |
| replace discrete convolution with SMPConv | CIFAR10 image classifier · CIFAR10 | accuracy: 92.9 -> 93.0 | Helped |
| replace RepLKNet-31B's large depth-wise kernels with SMPConv | RepLKNet-31B · ImageNet-1K | top-1 accuracy: 83.5 -> 83.8 | Helped |
| use SMPConv kernels instead of FlexConv kernels | CIFAR10 model with k=33 · CIFAR10 | training time (sec/epoch): 92.9 -> 40.1 | Helped |
| make radius learnable (ablation) | SMPConv with fixed point coordinates · CIFAR10 | accuracy: 90.92 -> 91.35 | Helped |
| make point coordinates learnable (ablation) | SMPConv with fixed radius · CIFAR10 | accuracy: 90.92 -> 92.47 | Helped |
| make both radius and point coordinates learnable (ablation) | SMPConv · CIFAR10 | accuracy: 90.92 -> 93.00 | Helped |
| initialize point locations with a wider spread, sigma=0.2, 0.3, or 0.5 instead of 0.05 (ablation) | SMPConv · CIFAR10 | accuracy: 93.00 -> 92.24; 91.84; 91.51 | Hurt |
| use a larger initial radius of 0.18, 0.24, or 0.3 instead of 0.12 (ablation) | SMPConv · CIFAR10 | accuracy: 93.00 -> 92.36; 92.06; 91.76 | Hurt |
| vary the number of moving points from 16 to 4, 8, 32, or 64 (ablation) | SMPConv · CIFAR10 | accuracy: 93.00 -> 92.56; 92.28; 92.84; 92.21 | Mixed |
Closest papers by meaning
- Continuity-Preserving Convolutional Autoencoders for Learning Continuous Latent Dynamical Models from Images (ICLR 2025 · similarity 0.85)
- KPConvX: Modernizing Kernel Point Convolution with Kernel Attention (CVPR 2024 · similarity 0.85)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.84)
- Spectral Convolutional Conditional Neural Process (NeurIPS 2025 · similarity 0.83)
- Recurrent Partial Kernel Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.83)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · 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.